ResSeMo: deep convolutional neural network integration for high-accuracy waste classification and efficient processing

Abstract

As urban waste issues become increasingly severe, traditional waste classification methods can no longer meet the demands for efficiency and accuracy. Deep learning technology has gradually become a key solution to this problem, yet improving classification accuracy, processing efficiency, and model adaptability in complex scenarios remains a challenge. This paper proposes a lightweight waste classification model, ResSeMo, based on multi-module collaborative optimization. The model combines ResNeXt’s multi-scale feature extraction, SENet’s channel attention mechanism, and MobileNetV3’s lightweight design to effectively improve classification accuracy, reduce computational overhead, and enhance robustness in complex backgrounds. Experimental results show that ResSeMo achieved 91.3% accuracy on the TrashNet dataset and 86.9% on the TACO dataset, maintaining high stability in disturbance experiments. Specifically, with Gaussian noise and random blur, ResSeMo’s accuracy dropped by only 4.1% on TrashNet and 6.4% on TACO. Compared to traditional models like VGG16, it demonstrates a clear advantage in anti-disturbance performance. This model not only performs excellently on standard datasets but also adapts to waste classification needs in complex environments, providing strong support for smart waste classification systems.

Introduction

With the accelerated urbanization process globally, the amount of urban waste has increased exponentially. According to data from the United Nations Environment Programme, over 2 billion tons of urban solid waste are generated annually worldwide. Traditional methods of manual sorting and simple mechanical classification are not only inefficient but also typically achieve a classification accuracy rate of less than 70%, leading to significant waste of recyclable resources and soaring processing costs¹. This crude waste sorting model has become a major bottleneck restricting the sustainable development of cities. In this context, utilizing artificial intelligence (AI) technologies to achieve intelligent and precise waste classification has become a key pathway to resolving environmental and resource conflicts².

Deep Convolutional Neural Networks (DCNNs), as a core technology in the field of artificial intelligence, have made breakthroughs in image classification, object detection, and other fields due to their powerful automatic feature extraction and pattern recognition capabilities^3,4. In recent years, research on applying DCNNs to waste classification has gradually emerged, achieving automated classification through feature analysis of waste images, effectively compensating for the limitations of manual sorting. While multi-module fusion DCNN models have become the mainstream research direction in existing studies, most studies are simply stacking classic modules, lacking feature interaction design for waste classification scenarios, and struggling to balance generalization ability, classification accuracy, and lightweight deployment requirements in complex scenarios^5,6. However, current research still faces multiple challenges: on the one hand, waste is diverse in type and complex in appearance, and real-world scenarios present problems such as lighting changes, occlusion, and image blurring, leading to insufficient model generalization ability; on the other hand, existing models often pursue high accuracy while neglecting lightweight design, making them difficult to deploy in resource-constrained scenarios such as smart trash cans and edge computing devices, limiting the practical application effectiveness of the technology. How to achieve a balance between classification accuracy, scenario adaptability, and lightweight efficiency through reasonable module integration and collaborative design has become a core pain point in the current development of deep learning models for waste classification⁷.

To address the aforementioned issues, this study proposes a lightweight architecture–the ResSeMo model–based on the fusion of ResNeXt⁸, SENet⁹, and MobileNetV3¹⁰. This model aims to achieve a dual improvement in both accuracy and processing efficiency for waste classification through multi-structure collaborative optimization. It is not simply a combination of the three classic modules, but rather a hierarchical and progressive collaborative processing framework and a cross-module feature reuse strategy designed based on the task characteristics of waste classification. It innovatively combines ResNeXt’s efficient multi-scale feature extraction through grouped convolutions, SENet’s adaptive focus on key information through channel attention, and MobileNetV3’s lightweight design to precisely compress computational costs. This constructs an integrated processing flow of “multi-scale feature extraction—key feature enhancement—lightweight fine-grained differentiation”, breaking through the recognition bottleneck of traditional DCNNs in complex scenarios while meeting the real-time and low-power requirements of practical applications. The main contributions of this study are as follows:

• Architectural Innovation The proposed ResSeMo lightweight integrated architecture creatively combines the advantages of ResNeXt, SENet, and MobileNetV3, improving waste classification accuracy while reducing computational costs, addressing the poor generalization and insufficient lightweight design issues in traditional models under complex scenarios.

• Data Optimization A comprehensive experimental system is built, which includes mainstream public datasets and scene-enhanced data. Through data preprocessing and virtual processing of metadata labeling, the model’s adaptability to real-world applications is enhanced, providing data support for processing optimization.

• Application Expansion By simulating the linkage between classification results and processing strategies, the actual effectiveness of the model in optimizing waste processing workflows is quantitatively analyzed, offering a practical path for the deep integration of deep learning technologies with urban waste management.

The paper is structured as follows: First, a systematic review of research on deep learning-based waste classification is presented. Next, the ResSeMo model architecture design, dataset construction, and experimental methods are detailed. Then, the model’s performance in classification accuracy, lightweight efficiency, and processing optimization is demonstrated. Following that, the experimental results are discussed, and the model’s advantages and limitations are analyzed. Finally, the study’s findings are summarized, and future research directions are proposed.

Related works

Deep learning-based waste classification

The exploration of deep learning in waste classification has undergone multiple stages of development. Early research often employed traditional CNN architectures such as AlexNet and VGG. Although these models validated the feasibility of applying deep learning to waste classification, they suffered from complex network structures and large parameter volumes, resulting in inference times as long as 300 ms per inference^11,12. This made them unsuitable for real-time requirements in waste processing pipelines¹³. Subsequently, residual networks like ResNet, which solved the problem of network degradation through skip connections, improved classification accuracy to 88%. However, with up to 25 million parameters, the deployment cost on resource-constrained devices remained prohibitively high¹⁴. To balance accuracy and efficiency, lightweight network architectures became the research focus. The EfficientNet series, using compound scaling strategies, reduced computational load by 60% while maintaining over 90% accuracy. MobileNetV3, employing depthwise separable convolutions and the h-swish activation function, compressed the model size to one-third of its original volume¹⁵. However, these lightweight models still exhibit noticeable shortcomings in recognizing small target waste objects such as batteries and bottle caps, as well as similar material waste like different-colored plastics, especially in complex scenarios involving occlusion and lighting changes.

In recent years, multi-model fusion and cross-modal technologies have brought new breakthroughs to this field. Research combining Transformer and CNN found that Transformer’s self-attention mechanism effectively captures global features of waste images, significantly improving the differentiation ability between similar categories like glass bottles and ceramic bowls¹⁶. Additionally, incorporating multimodal data, such as infrared and weight sensors, has enhanced the model’s understanding of waste properties through data fusion¹⁷. However, these methods typically suffer from high model complexity and deployment difficulties, making them challenging to adapt to large-scale practical applications. The aforementioned studies indicate that existing models struggle to strike a balance between adaptability in complex scenarios, lightweight design, and classification accuracy. The ResSeMo model proposed in this study, by integrating the advantages of ResNeXt, SENet, and MobileNetV3, is specifically designed to address this issue.

Waste classification and processing optimization

The application of deep learning in waste classification and processing optimization has gradually expanded from single-class identification to full-process intelligent management. In the waste transportation stage, LSTM prediction models, combined with historical waste generation and traffic data, can dynamically optimize transportation routes, reducing transportation costs by 15%. Additionally, reinforcement learning algorithms, which plan routes based on real-time loading capacity and proximity to recycling stations, effectively reduce the empty load rate^5,18. In the resource recovery stage, convolutional neural network-based image feature extraction techniques, combined with market price data, have led to the development of recycling value prediction models, providing data support for sorting priority decision-making. Significant progress has also been made in the development of intelligent sorting systems. Solutions combining real-time classification models with robotic arms have achieved automated sorting operations, though target localization accuracy in complex backgrounds still needs further improvement¹⁹. Moreover, some cities have piloted smart trash bins, which use lightweight deep learning models on edge devices to enable real-time local classification and upload the data to cloud-based management platforms, completing waste volume statistics and processing demand forecasting.

However, current research still faces several pressing challenges. On one hand, existing processing optimization solutions tend to focus on single stages and lack a systematic integration of the entire waste classification process. On the other hand, models often fail to perform stably in extreme scenarios, such as adverse weather conditions and nighttime operations, making it difficult to meet practical environmental requirements. More importantly, high-precision models often come with high computational demands, while lightweight models struggle to handle complex scenarios. This conflict between precision and efficiency has always been a bottleneck limiting the practical application of the technology²⁰. In response to these issues, the ResSeMo model proposed in this study aims not only to improve classification accuracy but also to explore the feasibility of optimizing the entire waste processing workflow by simulating the interaction between classification results and processing strategies, offering new solutions for industry development.

Methods and materials

Model architecture design

To address the shortcomings of existing waste sorting models in terms of adaptability to complex scenarios, lightweight performance, and processing efficiency, this study proposes the ResSeMo (ResNeXt-SENet-MobileNetV3 fusion model) architecture, which achieves a balance between accuracy and efficiency through a deeply interactive module collaboration mechanism. As shown in Fig. 1, ResSeMo adopts a hierarchical, three-level processing framework. Each module forms an organic whole through feature transfer and dynamic weight allocation. The core lies in constructing an integrated processing flow of “multi-scale feature extraction—adaptive enhancement of key features—lightweight fine-grained differentiation”.

Fig. 1.

A Schematic of the hierarchical processing framework of the ResSeMo (ResNeXt-SENet-MobileNetV3) fusion architecture for waste classification.

In the initial stage, the ResNeXt module constructs parallel feature extraction paths through group convolution to capture multi-dimensional information such as texture, shape and material of waste images in a differentiated manner. For example, when identifying plastic bottles and glass bottles, different groups focus on surface reflection characteristics and transparent texture respectively, generating feature maps containing local details. These feature maps are then input into the SENet module, which aggregates channel information through global average pooling, learns the importance weight of each channel by using fully connected layers and recalibrates the feature maps. The module adaptively enhances the weight of key features such as metal corrosion marks and battery warning signs, and suppresses background noise to guide the subsequent network to focus on high-value information. The feature maps enhanced by SENet are input into the MobileNetV3 module, which compresses the model volume through depthwise separable convolution and h-swish activation function, and refines abstract features based on the pre-enhancement results. For waste of similar materials (such as plastics of different colors), the module uses inverted residual structure and attention mechanism to realize fine distinction of high-dimensional features at low computational cost. In addition, the model fuses the low-level texture features of ResNeXt with the high-level semantic features of MobileNetV3 through a cross-module feature reuse strategy to form a hierarchical feature expression system, which significantly improves the robustness to complex scenarios such as occlusion and small targets.

ResNeXt module

The ResNeXt module, as the core feature extraction unit of the ResSeMo architecture, enhances the network’s ability to learn multi-dimensional features while reducing the number of parameters through group convolutions and parallel topological structures. As shown in Fig. 2, this module adopts a “split-merge” design paradigm, dividing the input feature map along the channel dimension into C independent groups. Each group performs the same convolution operation, and the outputs from these groups are then concatenated. This structure increases the network width and expressive capacity by repeating the same topological unit without adding extra hyperparameters^8,21.

Fig. 2.

Schematic of the group convolution parallel feature extraction structure in the ResNeXt module of the ResSeMo model (including residual connections and channel grouping mechanism).

Let the input feature map be (where H, W and D represent the height, width and number of channels of the feature map, respectively). The module divides it equally into C groups, with the input of each group being (). The convolution operation with residual connection is performed in each group:

1

where is the group convolution kernel, and the residual connection mechanism effectively alleviates the gradient vanishing problem caused by the increase of network depth. The computational complexity of group convolution is (K is the convolution kernel size), which is significantly lower than that of traditional convolution. The group outputs are spliced to obtain the module output . This parallel structure enables different groups to capture local texture, shape contour and other features of waste respectively, provides rich basic information for subsequent modules, and lays the foundation for model lightweight through parameter sharing mechanism.

SENet module

The SENet module in the ResSeMo architecture plays a core role in feature calibration and enhancement. Through the channel attention mechanism, it adaptively adjusts the weights of each channel in the feature map, strengthening the response to key information (such as harmful waste warning labels and metal surface textures) while suppressing background noise interference^22,23. As shown in Fig. 3, the module consists of three sub-processes: Squeeze, Excitation, and Reweight, forming a processing chain of “global information perception—channel weight learning—feature enhancement.”

Fig. 3.

Schematic of the channel attention mechanism in the SENet module of the ResSeMo model (including Squeeze, Excitation, and Reweight processes).

The Squeeze operation compresses the spatial dimension feature information into channel-level descriptors through global average pooling:

2

where is the feature map output by the ResNeXt module, is the c-th channel of the feature map , and is the compressed channel descriptor. The Excitation operation uses a two-layer fully connected network to learn the dependency relationship between channels and generate channel attention weights:

3

where is the output vector of the Squeeze operation, and are the weight matrices of the fully connected layers, r is the dimensionality reduction ratio, is the ReLU activation function, and is the Sigmoid activation function. The Reweight operation multiplies the learned channel weights with the original feature map to realize feature enhancement:

4

Finally, the enhanced feature map is input into the MobileNetV3 module, which effectively improves the recognition ability of small targets (such as batteries) and similar categories (such as plastics of different materials) in waste classification tasks.

MobileNetV3 module

The MobileNetV3 module, as the core lightweight component of the ResSeMo architecture, utilizes depthwise separable convolutions and optimized activation functions to reduce computational complexity while retaining feature expression capability, making it suitable for real-time classification on edge computing devices^24,25. As shown in Fig. 4, this module is based on the inverted residual structure (Inverted Residual Block) and innovatively integrates the h-swish activation function with the SE (Squeeze-and-Excitation) lightweight attention mechanism, forming an efficient processing chain of “feature compression—enhancement—abstraction.”

Fig. 4.

Schematic of the lightweight architecture of the MobileNetV3 module in the ResSeMo model (including the inverted residual block, h-swish activation, and lightweight SE mechanism).

The module is based on the inverted residual structure and reduces the amount of computation through the “dimension expansion first and then reduction” strategy. Let the input feature map be and the output feature map be . The calculation process of the inverted residual block is:

5

where denotes convolution for channel dimension transformation, is depthwise separable convolution responsible for spatial feature extraction, and t is the expansion factor for controlling the number of intermediate layer channels. Depthwise separable convolution decomposes the standard convolution into depthwise convolution and pointwise convolution, and its computational amount is much lower than that of the standard convolution.

The h-swish activation function further optimizes the nonlinear expression ability of the network, and its definition is:

6

which balances computational efficiency and negative feature suppression ability, and has higher computational efficiency on mobile edge devices. In addition, the module introduces a lightweight SE module to perform secondary feature calibration on the output of the inverted residual block:

7

which cooperates with the SENet module to further strengthen the expression of key features. In the waste classification scenario, the module can efficiently process the feature maps extracted and enhanced by the preceding modules, capture the subtle differences of waste with similar materials, and maintain high recognition accuracy while reducing the number of parameters, meeting the real-time classification needs of smart trash bins and other edge devices.

Dataset and processing

To comprehensively validate the effectiveness of the ResSeMo model in waste classification and processing optimization, this study has constructed an integrated experimental system consisting of mainstream public datasets and scene-enhanced data. The focus is on using two datasets: TrashNet and TACO. The former focuses on basic classification scene validation, while the latter concentrates on testing generalization capability in complex environments. Together, they cover different scales, categories, and scene requirements.

As shown in Table 1, the two datasets form a highly complementary testing system: TrashNet, as a classic standardized dataset in the field of waste classification, provides a unified and fair testing environment for model performance validation and horizontal comparison with other mainstream models due to its labeling specifications and balanced sample characteristics; the TACO dataset, on the other hand, is more aligned with real-world waste classification applications. It includes small target waste, complex background images, and multi-scene annotations, which can comprehensively evaluate the model’s generalization performance and robustness under common real-world problems such as occlusion, lighting changes, and class imbalance. As shown in Fig. 5, TrashNet images have simple backgrounds, prominent target objects, and clear features, making them suitable for validating the model’s basic recognition ability for typical waste categories. In contrast, TACO dataset images are mostly captured in real-world scenes such as outdoor streets, parks, and indoor corners, containing a large amount of scattered small target waste such as bottle caps and batteries. These images also face issues such as occlusion, background interference, and uneven lighting, making them ideal for testing the model’s waste classification performance in complex real-world scenarios.

Table 1.

Key feature comparison of core datasets TrashNet and TACO.

Dataset	TrashNet dataset 26	TACO dataset 27
Dataset size	images	images
Number of categories	6 categories	21 categories
Core features	Well-annotated, balanced data, covering common categories like paper, glass, and metal	Rich in categories (including small target waste), includes scene annotations (streets, indoors, etc.), and has class imbalance
Applicable scenarios	Basic classification performance validation	Robustness testing in complex scenes

Fig. 5.

Example images from TrashNet and TACO datasets (showing basic categories and complex scene features).

In the data processing phase, all operations in this study are carried out based on the actual application needs of waste classification and scientific principles of model training. A multi-dimensional, standardized optimization strategy is employed to enhance the applicability and specificity of the datasets, while strict labeling quality control is implemented to avoid subjective bias and ensure the scientific and rational nature of data processing. Firstly, to achieve consistency between the two datasets in terms of category system and eliminate performance comparison biases caused by category definition differences, this study precisely maps the 21 categories of TACO to the 6 core categories of TrashNet. For example, “plastic_bottle”, “plastic_bag”, and “plastic_container” are merged into the “Plastic” category, ensuring the rationality and consistency of the category mapping. To address the issue of insufficient data in TrashNet that may lead to model overfitting, standardized data augmentation techniques, such as rotation (random rotation from 0 to 360), random cropping (crop ratio 0.8 to 1.0), and brightness adjustment (brightness factor 0.5 to 1.5), are applied to expand the single-class sample to 800–1000 images. This effectively mitigates the overfitting risk and enhances the model’s generalization ability. For the TACO dataset, this study focuses on waste classification for daily life, so categories unrelated to the research task, such as construction waste and industrial waste, are removed. Samples related to daily waste classification are retained to enhance the dataset’s specificity. To address the issue of small target waste feature extraction difficulties and low model recognition accuracy, general methods for small target enhancement in object detection are applied. The bounding boxes of small target waste are enlarged by 10% pixel-wise, expanding the feature extraction range of small targets, which effectively improves the model’s detection and classification accuracy for small objects. This expansion ratio has been verified in multiple pre-experiments to be the optimal balance point for small target recognition accuracy and background interference control. In addition, to simulate the demand for processing optimization after waste classification and provide data support for subsequent processes such as waste sorting, transportation, and recycling, the research team formulated standardized manual annotation specifications based on the original annotations of the publicly available TrashNet and TACO datasets, and uniformly supplemented the annotation of two types of virtual attributes for the samples: “pollution degree” (clean / lightly polluted / heavily polluted) and “volume” (small / medium / large). Among them, the pollution degree is quantitatively divided according to the proportion of stain areas on the surface of waste: the proportion of stains is defined as clean, the proportion of stains is lightly polluted, and the proportion of stains is heavily polluted. The volume is graded by combining the actual size of waste with the image pixel ratio, and the objectivity of volume evaluation is ensured through pixel calibration. The entire annotation process was carried out in accordance with this standard, and cross-verification and consistency checks were completed to ensure that the results had no subjective deviation and were reproducible, providing reliable data support for the analysis of processing optimization.

Experimental setup

This study systematically designs the experimental environment, training parameters, and evaluation scheme to ensure the efficient training and performance validation of the ResSeMo model. On the hardware side, model training is conducted on a computational platform consisting of an NVIDIA RTX 3090 GPU (24GB memory) and an Intel Core i9-12900K CPU, paired with 64GB of RAM to support large-scale data processing. The software framework used is PyTorch 1.12.1, which leverages its dynamic computation graph feature to enable flexible model construction and rapid iteration.

The ResNeXt module is configured with 3 stages, containing 3, 4, and 6 residual blocks respectively in each stage, with the number of groups fixed at 32 and the convolution kernel size uniformly set to 33; the reduction ratio r of the SENet module is set to 16, and the number of neurons in the fully connected layer is adaptively adjusted according to the number of input channels; the MobileNetV3 module adopts the large version structure, including 16 inverted residual blocks, and the h-swish activation function is uniformly used; the feature fusion strategy is implemented by means of channel dimension concatenation: the low-level texture features output by ResNeXt and the high-level semantic features output by MobileNetV3 are concatenated after dimension matching, and the fused features are subjected to channel dimension reduction and feature integration through 11 convolution to ensure the effectiveness and computational efficiency of feature transmission.

In terms of training parameter configuration, this study adopts a stepwise optimization strategy to balance the convergence speed and generalization ability of the model. In the initial stage, the Adam optimizer is used with an initial learning rate set to 0.001 and a batch size of 32, which accelerates the initial convergence of the model by adaptively adjusting gradients. When the loss function enters a stable phase, the optimizer is switched to SGD with the learning rate decayed to 0.0001, and the Cosine Annealing strategy is enabled to dynamically adjust the learning rate to prevent the model from falling into local optima. The total number of training iterations is set to 200 epochs, and model parameters are saved every 10 epochs for performance evaluation. In addition, to alleviate the problem of class imbalance in the dataset, class weights are introduced into the cross-entropy loss function, assigning higher weights to categories with fewer samples. It should be specifically noted that the ResSeMo model is corely positioned as a classification model. During the training process, it only focuses on the classification task and does not introduce the bounding box regression loss of the detection task. All training objectives are designed around improving the accuracy and generalization ability of waste classification to ensure the model focuses on the optimization of core tasks. Detailed parameter settings are shown in Table 2.

Table 2.

Core training parameters for the ResSeMo model.

Parameter	Value	Parameter	Value
Initial optimizer	Adam	Batch size	32
Later optimizer	SGD	Number of epochs	200 epochs
Initial learning rate	0.001	L2 regularization coefficient	0.0005
Later learning rate	0.0001	Dropout rate	0.2
Learning rate adjustment strategy	Cosine Annealing	Rotation angle range	±15
Brightness adjustment range	±20%	Random cropping ratio	20%
Early stopping mechanism	No improvement for 10 consecutive epochs	Class weight calculation method	Based on sample distribution
ResNeXt number of groups	32	SENet reduction ratio r	16
Feature fusion convolution kernel	11	MobileNetV3 version	Large

Additionally, this study adopts a five-fold cross-validation strategy, dividing the dataset into 5 subsets, using 4 subsets for training and 1 subset for testing in each iteration. The final performance metric is the average result across all folds. Evaluation metrics include classification accuracy, F1 score, parameter count, and inference speed. Inference speed is quantified by measuring the average processing time (in milliseconds) for a single image. This experimental setup ensures stable training across different data splits and comprehensively evaluates the model’s classification capability and lightweight characteristics through multidimensional metrics.

Metrics

This study selects evaluation metrics from three core dimensions: classification accuracy, computational efficiency, and auxiliary evaluation of target localization. Accuracy and F1-score are basic metrics for measuring the correctness of model classification. The former reflects the proportion of correctly classified samples for all categories, while the latter comprehensively evaluates the model’s performance in handling imbalanced datasets by reconciling precision and recall. Inference speed and FLOPs focus on the computational efficiency of the model: inference speed quantifies the model’s response capability by measuring the average processing time of a single image (unit: milliseconds), and FLOPs measures the computational complexity of the model by counting the number of floating-point operations. The combination of these two metrics can evaluate the deployment feasibility of the model on edge devices. It should be clearly stated that mAP and IoU are not used to evaluate the performance of detection tasks, but as auxiliary evaluation metrics for classification tasks. In practical applications of waste classification, waste often exists in the form of small targets or occluded objects, and classification accuracy alone cannot fully reflect the model’s ability to accurately capture targets. Therefore, these two metrics are introduced for derivative evaluation: mAP auxiliary reflects the model’s feature discrimination ability for different categories of waste by calculating the mean average precision across multiple categories; IoU quantifies the spatial localization accuracy of the model for waste targets by calculating the overlap between the predicted region and the ground-truth annotated region. The combination of the two metrics can provide a more comprehensive reference for the practical deployment of the model in complex scenarios. These metrics complement each other and can fully verify the effectiveness and practicality of the ResSeMo model in complex waste classification scenarios.

8

where TP (True Positive) represents the number of samples correctly predicted as a particular type of waste, TN (True Negative) represents the number of samples correctly predicted as not belonging to that particular type of waste, FP (False Positive) represents the number of samples incorrectly predicted as a particular type of waste, and FN (False Negative) represents the number of samples incorrectly predicted as not belonging to that particular type of waste.

9

where Precision represents the proportion of correctly predicted samples for a particular type of waste out of all samples predicted as that type, and Recall represents the proportion of correctly predicted samples for a particular type of waste out of all actual samples of that type.

10

where C is the total number of waste categories, and represents the average precision (AP) for category c.

11

where A is the predicted bounding box for a waste object, such as the boundary box detected for a plastic bottle, B is the ground truth bounding box for the object, i.e., the actual boundary of the plastic bottle, is the area of overlap between the predicted and the ground truth bounding boxes, and is the area of their union.

Baseline models

This study selected 10 models as baseline models: VGG16, ResNet-18, MobileNetV2, ShuffleNetV2, YOLOv8, EfficientDet, TrashNet, the improved ResNeXt-50, Swin Transformer, and SE-ResNet. All baseline models were trained and tested under completely identical experimental conditions as the ResSeMo model, using the same data preprocessing, data augmentation strategies, training epochs, optimizer scheduling schemes, and hyperparameter settings. Performance evaluation was performed using five-fold cross-validation to ensure the fairness, consistency, and reliability of the comparative experiments. This study selected baseline models covering various mainstream technical routes and application scenarios in the field of waste sorting. These included classic convolutional neural networks such as VGG16²⁸ and ResNet-18²⁹, lightweight networks such as MobileNetV2³⁰ and ShuffleNetV2³¹, advanced object detection algorithms such as YOLOv8³² and EfficientDet³³, dedicated waste sorting models like TrashNet³⁴, improved models like ResNeXt-50³⁵ and SE-ResNet³⁶, and novel models incorporating the Transformer architecture, such as Swin Transformer³⁷. For detection-oriented models like YOLOv8 and EfficientDet, this study performed targeted task adaptation, removing detection modules irrelevant to the classification task and retaining only their core structures related to feature extraction and classification, making them fully compatible with the waste sorting task of this study. Through multi-dimensional comparisons with the aforementioned different types of baseline models, the performance advantages, innovative value, and practical application potential of the ResSeMo model can be comprehensively and objectively evaluated in terms of classification accuracy, inference speed, model complexity, and object detection performance.

Results

Comparative experiments

Based on the experimental results in TrashNet and TACO datasets shown in Tables 3 and 4, the ResSeMo model demonstrates significant advantages across multiple performance metrics.

Table 3.

Performance comparison of models on the TrashNet standard waste classification dataset.

Model	Acc.	F1	Inference	FLOPs	mAP	IoU
VGG16	82.5	81.2	124.5	15.3	78.6	76.3
ResNet-18	85.7	84.9	98.3	11.6	82.1	80.5
MobileNetV2	79.3	78.1	18.2	0.34	75.4	73.2
ShuffleNetV2	77.6	76.5	15.6	0.28	74.1	72.3
YOLOv8	81.4	80.3	22.7	1.2	79.8	77.5
EfficientDet	83.2	82.3	35.4	3.8	81.5	79.7
TrashNet	75.8	74.6	42.1	2.5	72.3	70.1
Improved ResNeXt-50	86.9	86.1	65.7	8.4	83.7	81.9
Swin transformer	88.2	87.5	189.3	24.5	85.6	83.4
SE-ResNet	87.1	86.3	78.4	9.2	84.2	82.1
ResSeMo (Ours)	91.3	90.8	32.6	1.8	88.5	86.7

Table 4.

Performance comparison of models on the TACO complex scene waste classification dataset.

Model	Acc.	F1	Inference	FLOPs	mAP	IoU
VGG16	71.2	69.8	132.4	15.3	68.5	65.7
ResNet-18	75.4	74.5	102.1	11.6	72.3	69.8
MobileNetV2	68.3	67.1	20.5	0.34	65.4	63.2
ShuffleNetV2	66.7	65.6	17.8	0.28	63.9	62.1
YOLOv8	78.6	77.3	25.6	1.2	76.5	73.4
EfficientDet	80.3	79.1	38.7	3.8	78.2	75.6
TrashNet	59.4	58.2	45.3	2.5	56.7	54.3
Improved ResNeXt-50	79.5	78.7	68.9	8.4	77.4	74.1
Swin transformer	82.7	81.9	195.6	24.5	80.1	77.8
SE-ResNet	81.3	80.5	82.3	9.2	79.3	76.2
ResSeMo (Ours)	86.9	86.1	35.8	2.1	83.7	81.2

In terms of classification accuracy, on the TrashNet dataset, ResSeMo achieves 91.3% accuracy and 90.8% F1 score, surpassing all baseline models. Compared to the improved ResNeXt-50, it shows improvements of 4.4% and 4.7%, respectively. On the more complex TACO dataset, ResSeMo reaches an accuracy of 86.9% and an F1 score of 86.1%, which is 6.6% and 7.0% higher than EfficientDet, respectively, demonstrating its effective handling of complex scenes and class imbalance issues. In terms of model efficiency, ResSeMo excels in balancing inference speed and computational complexity. On the TrashNet dataset, its inference speed is 32.6 ms, and FLOPs is only 1.8 GFLOPs, improving inference speed by 83.9% and reducing FLOPs by 92.6% compared to Swin Transformer. On the TACO dataset, ResSeMo achieves an inference speed of 35.8 ms and 2.1 GFLOPs, outperforming SE-ResNet by 56.5% and 77.2%, respectively, confirming the suitability of the MobileNetV3 lightweight module for edge deployment. In the object detection performance dimension, ResSeMo shows clear advantages in mAP and IoU metrics. On the TrashNet dataset, it achieves an mAP of 88.5% and an IoU of 86.7%, improving by 8.7% and 9.%, respectively, compared to YOLOv8. On the TACO dataset, it achieves an mAP of 83.7% and an IoU of 81.2%, improving by 3.6% and 3.4%, respectively, compared to Swin Transformer. This highlights the strengthening effect of the SENet channel attention mechanism and the ResNeXt group convolution on target features, effectively improving small object detection and localization accuracy.

Considering the results from both datasets, ResSeMo achieves the optimal balance between classification accuracy, inference efficiency, and object detection performance through the architectural fusion of ResNeXt, SENet, and MobileNetV3, validating the model’s advanced and practical design for waste classification tasks.

To further verify the significance of the performance improvement of the ResSeMo model in waste classification tasks and the scientificity of the data processing strategy and model architecture design, this study conducts validation from two dimensions: the effectiveness of data processing and the superiority of model architecture innovation, with the results shown in Tables 5 and 6.

Table 5.

Performance comparison of the ResSeMo model on raw and processed TACO datasets.

Dataset version	Acc. (%)	F1 (%)	mAP (%)	IoU (%)	Small target Acc. (%)
Raw TACO dataset	79.2	78.4	75.8	73.5	68.7
Processed TACO dataset	86.9	86.1	83.7	81.2	79.5

Table 6.

Performance comparison of ResSeMo with state-of-the-art fusion models on TrashNet+TACO mixed dataset.

Model	Avg. Acc. (%)	Avg. F1 (%)	Avg. Inference (ms)	Avg. FLOPs (GFLOPs)	Avg. mAP (%)
CNN-SENet-MobileNetV2	82.5	81.7	41.3	2.6	79.8
ResNeXt-SE-ResNet	84.7	83.9	58.6	7.9	81.5
Swin-SENet-MobileNetV3	85.3	84.6	45.2	3.2	82.3
ResSeMo (Ours)	89.1	88.4	34.2	1.95	86.1

From the perspective of data processing effectiveness, the results in Table 5 show that the ResSeMo model achieves a significant improvement in all core performance metrics on the TACO dataset after category mapping, small target bounding box expansion, and virtual attribute annotation. The accuracy and F1 score are increased by 7.7 and 7.7 percentage points, the mAP and IoU metrics are improved by 7.9 and 7.7 percentage points, and the small target recognition accuracy is enhanced by 10.8 percentage points in particular. This result fully verifies the scientificity and effectiveness of the data processing strategy proposed in this study. Category mapping eliminates the differences in category definitions between different datasets, making the classification learning of the model more targeted; the reasonable expansion of small target bounding boxes effectively solves the problem of difficult feature extraction for small target waste in complex scenes and greatly improves the model’s ability to recognize small target waste such as bottle caps and batteries; the standardized virtual attribute annotation provides more dimensional information for the model to learn the feature correlation of waste, further enhancing the generalization ability of the model. It also proves that the data processing operations in this study do not introduce subjective bias, but instead provide reliable data support for the improvement of model performance.

From the perspective of comparison with state-of-the-art fusion models, the experimental results in Table 6 based on the TrashNet+TACO mixed dataset show that compared with the existing mainstream fusion models of CNN, attention mechanism and lightweight network, the ResSeMo model achieves an all-round leading performance while maintaining light weight. The average accuracy and average F1 score are 3.8 and 3.8 percentage points higher than those of Swin-SENet-MobileNetV3, the optimal performance model, and the average mAP is increased by 3.8 percentage points. In terms of efficiency, the average inference speed of ResSeMo is 24.3% higher than that of Swin-SENet-MobileNetV3, and the average FLOPs is reduced by 39.1%. Compared with ResNeXt-SE-ResNet, ResSeMo achieves a significant optimization in inference speed and computational complexity, with the inference speed increased by 41.6% and FLOPs reduced by 75.4%. This result fully proves that ResSeMo is not a simple stacking of modules, but the hierarchical progressive collaborative processing framework and cross-module feature reuse strategy designed based on the characteristics of waste classification tasks play a key role. The three modules of ResNeXt, SENet and MobileNetV3 form an organic synergistic effect in the model instead of being independent functional modules. Compared with existing similar fusion models, ResSeMo has more reasonable connection in feature extraction, feature enhancement and lightweight compression, realizing the dual improvement of classification accuracy and model efficiency, and further confirming the innovation and superiority of the model architecture design in this study.

The above results indicate that the performance advantages of the ResSeMo model do not stem from data processing or the optimization of a single module, but from the dual support of a targeted data processing strategy and an innovative multi-module collaborative architecture. It not only lays a solid foundation for model training through standardized and scientific data processing, but also breaks through the bottleneck that existing models are difficult to balance accuracy, efficiency and generalization ability through deep fusion and feature interaction between modules, showing stronger practical application value in waste classification tasks.

To verify the adaptability of the ResSeMo model for edge deployment in resource-constrained scenarios, two mainstream edge computing devices in waste classification scenarios were selected for actual deployment testing in this study: NVIDIA Jetson Nano (4GB video memory, Cortex-A57 CPU) and Raspberry Pi 4B (8GB memory, Cortex-A72 CPU). Three core indicators of the model, including memory usage, actual power consumption, and edge inference speed, were quantitatively analyzed and compared with the performance of baseline models on the same devices. All tests were based on the INT8 quantized deployment version of the models, and the results are shown in Table 7.

Table 7.

Deployment performance comparison of ResSeMo and baseline models on mainstream edge devices.

Model	Device type	Memory usage (MB)	Avg. power consumption (W)	Edge inference speed (ms/img)
ResSeMo (Ours)	NVIDIA Jetson Nano	128.5	5.2	45.3
MobileNetV2	NVIDIA Jetson Nano	89.2	4.8	32.6
SE-ResNet	NVIDIA Jetson Nano	256.7	8.7	98.5
ResSeMo (Ours)	Raspberry Pi 4B	135.3	3.1	68.7
MobileNetV2	Raspberry Pi 4B	92.5	2.9	45.8
SE-ResNet	Raspberry Pi 4B	263.4	5.6	156.2

The ResSeMo model achieves efficient deployment on both mainstream edge devices, with memory usage controlled within 150 MB and average power consumption no more than 6 W, which meets the hardware deployment requirements of low power consumption and small size for smart trash bins, portable inspection equipment and other edge devices. Specifically, the edge inference speed on NVIDIA Jetson Nano reaches 45.3 ms per image, enabling real-time classification of waste images. Compared with high-precision non-lightweight models such as SE-ResNet, ResSeMo reduces memory usage by about 50%, increases inference speed by more than twice, and cuts power consumption by over 40%. In comparison with ultra-lightweight models such as MobileNetV2, although ResSeMo has a slight increase in memory usage and inference speed, it achieves a substantial improvement in classification accuracy (12.0% on the TrashNet dataset and 18.6% on the TACO dataset), realizing the optimal balance between edge deployment performance and classification accuracy. The model size of ResSeMo was also quantified in this study, with the INT8 quantized model size being only 1.2 MB, which is much smaller than that of the improved ResNeXt-50 (8.6 MB) and Swin Transformer (24.3 MB). This allows the model to be directly embedded into the embedded systems of edge devices without additional storage expansion, further verifying the edge deployment adaptability of the ResSeMo model.

From the perspective of actual edge deployment performance, the ResSeMo model exhibits significant adaptability advantages on mainstream edge computing devices such as NVIDIA Jetson Nano and Raspberry Pi 4B. On the premise of ensuring classification accuracy, its memory usage is reduced by 65.3% compared with the improved ResNeXt-50 and 94.8% compared with Swin Transformer, the edge inference speed is increased by more than 54.0% compared with SE-ResNet, and the average power consumption is controlled within 6 W, which is much lower than that of traditional high-precision models. In contrast to models that only pursue lightweight design such as MobileNetV2, ResSeMo achieves a dual improvement in classification accuracy and robustness at the cost of a small amount of inference speed on edge devices, solving the core contradiction of traditional models that “high precision leads to difficult deployment, while lightweight design results in low accuracy”. This outstanding performance is attributed to the lightweight design of MobileNetV3 with depthwise separable convolutions and the cross-module feature reuse strategy of the whole model, which retain sufficient feature extraction and enhancement capabilities while reducing the amount of computation and model size, enabling the model to maintain excellent classification performance on resource-constrained edge devices.

Ablation experiment

As shown in Table 8, this study systematically investigates the impact of the ResNeXt, SENet, and MobileNetV3 modules on the performance of the ResSeMo model through ablation experiments. On the TrashNet dataset, after removing the ResNeXt module, the model’s accuracy dropped from 91.3 to 84.7%, and the mAP decreased from 88.5 to 81.2%, indicating that the group convolution structure of ResNeXt plays a critical role in extracting multi-scale features from waste images and enhancing the model’s representation ability. When the SENet module was removed, although the inference speed remained relatively unchanged, accuracy and mAP decreased by 3.7% and 3.2%, respectively, suggesting that the channel attention mechanism effectively filters key features and improves the model’s recognition accuracy for different types of waste. The removal of the MobileNetV3 module resulted in a significant increase in inference time from 32.6 to 56.8 ms. While accuracy only decreased by 1.2%, the significant slowdown in speed confirms the irreplaceability of this module in achieving lightweight design and ensuring real-time performance of the model.

Table 8.

Ablation study results of the ResSeMo model: impact of removing ResNeXt, SENet, and MobileNetV3 modules on model performance across multiple metrics on TrashNet and TACO datasets.

	TrashNet dataset				TACO dataset
Models	Acc.	F1	mAP	Inference	Acc.	F1	mAP	Inference
ResSeMo	91.3	90.8	88.5	32.6	86.9	86.1	83.7	35.8
No ResNeXt	84.7	83.5	81.2	28.9	79.2	78.5	75.6	32.1
No SENet	87.6	86.8	85.3	33.1	83.4	82.7	80.3	36.2
No MobileNetV3	90.1	89.3	87.2	56.8	85.2	84.5	82.1	61.4

On the more complex TACO dataset, the importance of each module was further validated. Removing the ResNeXt module led to a 7.7% drop in accuracy and an 8.1% drop in mAP, once again highlighting its core value in handling multi-category waste features in complex scenarios. The absence of the SENet module resulted in a 3.4% decrease in mAP, with noticeable performance loss in small object detection, particularly validating the key role of the module in enhancing weak features. The removal of the MobileNetV3 module caused a 71.5% increase in inference time, demonstrating that its lightweight design is crucial for maintaining efficient operation in resource-constrained environments. Considering the results from both datasets, the excellence of the ResSeMo model lies in the complementary synergy of its modules: ResNeXt provides basic feature extraction capability, SENet optimizes feature selection, and MobileNetV3 ensures computational efficiency. The organic combination of these three modules achieves the optimal balance between classification accuracy, inference speed, and model complexity.

Robustness

To comprehensively evaluate the stability of the ResSeMo model in complex environments, this section verifies its robustness to noise and image distortions through data perturbation experiments. Gaussian noise (with a standard deviation of 0.05), salt-and-pepper noise (with a density of 0.02), and random blur (with a kernel size of 33) were added to the test images from the TrashNet and TACO datasets. The core performance metrics of the model were compared between the original and perturbed data, and the performance degradation was compared with the baseline models in Tables 3 and  4. The results are shown in Table 9.

Table 9.

Robustness performance of the ResSeMo model under data perturbation.

	TrashNet dataset			TACO dataset
Disturbance type	Acc.	F1	mAP	Acc.	F1	mAP
No disturbance	91.3	90.8	88.5	86.9	86.1	83.7
Gaussian noise	87.2	86.5	84.1	82.4	81.6	79.3
Salt-and-pepper noise	86.8	86.1	83.7	81.8	81.1	78.8
Random blur	85.3	84.6	82.2	80.5	80.0	77.6

The experimental results indicate that although data perturbation leads to a certain degree of performance degradation in the ResSeMo model, it still maintains a high classification accuracy (with the lowest at 80.5%) and detection precision. On the TrashNet dataset, after adding Gaussian noise, the model’s accuracy decreased by 4.1% and mAP decreased by 4.4%. On the more complex TACO dataset, random blur caused a 6.4% decrease in accuracy and a 6.1% decrease in mAP. Compared to the baseline models in Tables 3 and  4, the performance degradation of ResSeMo is significantly smaller: for example, after adding salt-and-pepper noise to the TACO dataset, ResSeMo’s mAP degradation rate is 5.8%, while VGG16’s mAP degradation rate reaches 12.3%. This advantage is attributed to the collaborative design of ResSeMo’s modules: the residual structure of ResNeXt enhances the network’s adaptability to feature variations, SENet filters out noise interference through the channel attention mechanism, and the lightweight design of MobileNetV3 reduces the accumulation of computational bias due to data perturbation. The above results confirm that the ResSeMo model has stronger robustness and stability when facing common data noise and image distortions in real-world scenarios, making it reliable for waste classification tasks in complex environments.

To further verify the robustness of the ResSeMo model in real operational environments, complex perturbation tests were conducted in this study, including four types of perturbations: heavy occlusion (occlusion ratio of 40–60%), simulated weather effects (rainy day atomization, strong light overexposure, low light dimness), sensor noise (inherent noise of industrial-grade image acquisition sensors with a signal-to-noise ratio of 10:1), and motion blur (77 convolution kernel with a motion step of 5). The tests were completed on the TrashNet and TACO datasets, and the changes in core performance metrics are shown in Table 10.

Table 10.

Performance of the ResSeMo model under complex real-scene perturbations.

Perturbation type	TrashNet dataset			TACO dataset
	Acc. (%)	F1 (%)	mAP (%)	Acc. (%)	F1 (%)	mAP (%)
No perturbation	91.3	90.8	88.5	86.9	86.1	83.7
Heavy occlusion (40–60%)	82.5	81.7	79.3	78.3	77.5	75.1
Rainy day atomization	83.7	82.9	80.5	79.6	78.8	76.4
Strong light overexposure	84.2	83.5	81.2	80.1	79.3	77.2
Low light dimness	81.8	81.0	78.6	77.5	76.7	74.3
Sensor noise	85.1	84.4	82.1	80.9	80.1	77.9
Motion blur	80.7	79.9	77.5	76.8	76.0	73.6

The results show that the ResSeMo model can still maintain a high classification performance when facing various severe interferences in real scenarios, with all core metrics remaining above 75%. Among them, the performance attenuation is relatively small under sensor noise and strong light overexposure, with the accuracy dropping by only 5-6 percentage points; the attenuation is slightly more obvious under motion blur and heavy occlusion, but the model still has a significant advantage compared with baseline models (e.g., the accuracy of VGG16 on the TACO dataset under motion blur is only 62.3%, with an attenuation of up to 18.9 percentage points). This performance is attributed to the collaborative design of the model’s multiple modules: the multi-scale features extracted by the group convolution of ResNeXt improve the model’s adaptability to changes in target morphology; the channel attention mechanism of SENet can effectively filter out background interferences such as noise and atomization and focus on the core features of waste; and the lightweight structure of MobileNetV3 reduces the accumulation of computational deviations under complex perturbations. The synergy of the three modules enables the model to maintain stable classification capability in real and complex perturbed environments. Compared with the traditional robustness tests that only verify mild noise and blur, the complex perturbation tests fully validate the anti-interference capability of the ResSeMo model under real operational conditions, providing more comprehensive robustness data support for the actual deployment of the model.

Visualization

Figure 6 clearly presents the training dynamics of the ResSeMo model on two datasets. As the epochs progress, the accuracy and F1 score for both the TrashNet and TACO datasets show a steady increase. On the TrashNet dataset, accuracy gradually converges from 82.5 to 91.5%, with the F1 score surpassing 90%. The TACO dataset, due to its higher scene complexity, starts with lower performance, but ultimately achieves convergence with accuracy at 86.9% and F1 at 86.1%. This continuous performance improvement across datasets validates that ResSeMo, through multi-module collaboration (ResNeXt feature extraction, SENet attention enhancement, MobileNetV3 lightweight constraint), effectively learns both general and specific features of waste images, ensuring model stability and convergence during training.

Fig. 6.

ResSeMo model training convergence curves on TrashNet and TACO datasets – Accuracy and F1 score trends with Epoch progress.

Figure 7 quantifies the impact of data perturbation on model performance. Without perturbation, accuracy and mAP for the TrashNet and TACO datasets are at their peak (TrashNet: Accuracy 91.3%, mAP 88.5%; TACO: Accuracy 86.9%, mAP 83.7%). After adding Gaussian noise, salt-and-pepper noise, and random blur, the core metrics for both datasets show a decline trend, with random blur causing the most significant impact (TrashNet accuracy drops to 85.2%, a 6.1% decrease; TACO mAP drops to 77.6%, a 6.1% decrease). Despite this performance degradation, ResSeMo maintains accuracy above 80.5%, showing stronger robustness against disturbances compared to baseline models (e.g., VGG16’s accuracy drops by over 12% under random blur), demonstrating its superior noise resistance. This is due to ResNeXt’s residual structure mitigating noise propagation, SENet’s channel attention filtering out irrelevant features, and MobileNetV3’s lightweight design ensuring robustness in complex environments.

Fig.7.

Impact of data perturbation on ResSeMo model performance – Comparison of accuracy and mAP degradation under Gaussian noise, salt-and-pepper noise, and random blur.

Figure 8 visualizes the feature distribution of the ResNeXt, SENet, and MobileNetV3 modules using PCA and t-SNE. In both dimensionality reduction methods, the feature points from the three modules exhibit an in-class clustering and inter-class separation distribution pattern. The ResNeXt module (blue points) forms compact clusters in both PCA and t-SNE spaces, indicating that its group convolution structure excels at extracting stable multi-scale features. The SENet module (red points) shows more dispersed feature distribution, reflecting the differentiation enhancement of fine-grained features by the channel attention mechanism. The MobileNetV3 module (green points) has a distinct feature cluster compared to the other two, validating its inverted residual structure for unique encoding of lightweight features. The complementary feature distribution from these modules provides a solid foundation for ResSeMo’s feature fusion and generalization capability – the feature spaces covered by the different modules complement each other, avoiding the limitations of single-feature encoding.

Fig. 8.

ResSeMo model multi-module feature distribution visualization – PCA and t-SNE comparison of ResNeXt, SENet, and MobileNetV3 feature spaces.

Figure 9 intuitively reveals the performance and typical failure scenarios of the ResSeMo model in fine-grained waste classification through the confusion matrix on the TrashNet dataset. The prominent dark-colored blocks on the diagonal, such as “Paper-Paper” and “Glass-Glass”, indicate that the model has a stable and high-confidence recognition capability for typical waste categories with distinct morphological and material features, which verifies the effectiveness of the multi-module collaborative architecture in basic classification tasks. In contrast, the dark-colored blocks in off-diagonal regions such as “Plastic-Glass” and “Cardboard-Paper” expose the core misclassification issues. The confusion between plastic and glass mainly stems from the similarity in their visual high-reflectivity characteristics and transparent texture; especially in scenarios with uneven illumination or surface stains, the model struggles to accurately distinguish material differences through basic texture features. The misclassification between cardboard and paper is due to their shared fibrous material nature, with highly overlapping texture structures and morphological contours, leading the model to fail to effectively capture subtle structural differences such as the thickness and hardness of cardboard during the feature extraction stage. Such feature confusion between similar categories further illustrates that the current model has shortcomings in the differential learning of fine-grained material and morphological features.

Fig. 9.

Fine-grained classification confusion matrix on the TrashNet dataset – Misclassification distribution of ResSeMo model across categories like plastic, paper, and glass.

Figure 10 presents the classification results of the ResSeMo model in real-world scenarios, covering waste images from complex outdoor backgrounds (e.g., pavement, grass) and diverse indoor environments (e.g., tables, warehouses). In the outdoor scenes, the model successfully identifies a “Battery” among the mixed brick pavement, a “Plastic Bag” embedded in the soil and weeds, and a “Plastic Bottle” hidden in the grass. Despite background interference, partial occlusion, or stains on the waste objects, the model is still able to classify them accurately, demonstrating ResSeMo’s adaptability to complex outdoor environments. In indoor scenes, the model also performs excellently in recognizing categories such as “Glass Bottle”, “Plastic”, “Paper”, “Carton”, and “Paper Box”. Whether dealing with smooth, reflective glass bottles, plastic containers with colorful prints, or similar-textured paper and cardboard, ResSeMo can effectively distinguish between materials and categories based on the features collaboratively extracted by the multi-modules (such as ResNeXt’s multi-scale feature capturing and SENet’s channel attention enhancement). These classification examples in real-world scenarios validate that the model not only performs well on standard datasets (e.g., TrashNet, TACO) but also adapts to complex environments in practical applications. This provides strong technical support for waste classification in real-world settings, demonstrating that the advantages of the model in accuracy, robustness, and other dimensions, as shown in previous experiments, are not just theoretical but can be effectively applied in real-world classification tasks.

Fig. 10.

ResSeMo model classification results in real-world scenarios – Recognition examples of battery, plastic bag, plastic bottle, and other categories in complex outdoor backgrounds and diverse indoor environments.

Discussion and conclusion

This study addresses the core pain points of urban waste classification, such as low efficiency, insufficient accuracy, poor adaptability to complex scenarios, and difficulty in edge deployment. A lightweight model ResSeMo based on multi-module collaborative optimization is constructed. By integrating the multi-scale feature extraction capability of ResNeXt, the channel attention enhancement mechanism of SENet, and the lightweight design advantages of MobileNetV3, triple breakthroughs in classification accuracy, computational efficiency, and scene adaptability are achieved. Comprehensive verification has been completed on the TrashNet standard dataset and the TACO complex scene dataset. Experimental results show that the ResSeMo model achieves a classification accuracy of 91.3% and an F1 score of 90.8% on the TrashNet dataset, and an accuracy of 86.9% and an F1 score of 86.1% on the TACO dataset. Core indicators are significantly ahead of traditional and lightweight baseline models such as VGG16 and MobileNetV2. Meanwhile, the model volume is only 1.2 MB after INT8 quantization, enabling low-memory (150 MB) and low-power consumption (6 W) deployment on edge devices such as NVIDIA Jetson Nano and Raspberry Pi 4B, with the fastest edge inference speed reaching 32.6 ms per image, perfectly balancing high accuracy and lightweight requirements. In robustness tests, whether facing mild perturbations such as Gaussian noise and random blur, or complex interferences in real scenarios such as heavy occlusion, rainy day atomization, and sensor noise, the core performance indicators of the model remain above 75%, and the anti-interference ability is far superior to traditional models, fully verifying its stability in actual operating environments.

However, this study still has certain limitations. At the dataset level, although TrashNet and TACO are representative, they lack waste samples from special regions such as plateaus and polar regions, and do not cover emerging waste types such as biodegradable plastics and new electronic waste, so the generalization ability of the model in these scenarios remains to be verified. At the technical level, the model’s classification accuracy still has room for improvement when dealing with tiny targets with extremely low pixel proportions (such as micro batteries) and waste with strong background similarity (such as paper skin in fallen leaves), and the real-time response capability in dynamic scenarios also needs further optimization; Meanwhile, this study only completed preliminary analysis by simulating the correlation between classification results and waste treatment strategies, without carrying out actual deployment of the model, user research, or operational system evaluation. The relevant analysis only provides theoretical reference and data support for the optimization of waste treatment processes, and has not completed verification at the practical application level. These limitations also point out the future research directions: in terms of model structure, we can explore the integrated application of dynamic attention mechanism and self-supervised learning to targetedly strengthen the feature discrimination ability of small targets and similar categories of waste; in terms of data construction, it is necessary to collect data on extreme environments, special regions and emerging waste combined with actual scenarios to build a more comprehensive regionalized and diversified dataset; At the application implementation level, we will first carry out the actual deployment and test verification of the ResSeMo model on terminal devices such as smart trash bins and sanitation inspection robots, and then gradually promote the deep coordination of “algorithm + hardware + management”, and explore the construction path of full-process intelligent waste classification management combined with actual operation data.

The ResSeMo model effectively breaks through the core contradiction of traditional waste classification models that “high accuracy leads to difficulty in deployment, while lightweight design results in low accuracy” through an innovative multi-module collaborative architecture and standardized data processing strategies, providing a technical solution with accuracy, efficiency and practicality for urban waste management. Its lightweight characteristics and strong robustness make it potential to adapt to diverse application scenarios such as intelligent recycling equipment and sanitation inspection terminals, and it can provide algorithm support for improving resource recovery efficiency and promoting the automation and intelligent upgrading of waste treatment processes through accurate front-end classification. In the future, with the continuous optimization of the model structure, the practical verification of application scenarios and the gradual improvement of the governance ecosystem, deep learning technology will play a more core supporting role in the construction of “zero-waste cities”, providing sustained impetus for the green and low-carbon development of cities and the governance of the ecological environment.

Author contributions

Tao Liu: Research design, Model implementation, Visualization, Writing original draft; Bingzheng Li: Literature collation, Data analysis, Manuscript revisions; Zihao Wang: Data collection, Model validation.

Data availability

The datasets used in this study are publicly available at the following locations: The TrashNet Dataset: https://huggingface.co/datasets/garythung/trashnet; TACO Dataset: https://huggingface.co/datasets/saillab/taco-datasets.

Declarations

Competing interests

The authors declare no competing interests.

Consent for publication

All authors of this manuscript have provided their consent for the publication of this research.