GroupCeptionNet: a lightweight model for classifying chili seed germination with macro images

Abstract

Chili seed germination potential is a significant element influencing planting yield. Many researchers focus on feature extraction analysis methods to detect undesirable seed germination characteristics and improve classification accuracy. In deep learning, data collection and model complexity often limit practical application in production environments. This work includes a dataset of macroscopic images of chili seeds. This paper proposes a lightweight model GroupCeptionNet to process macroscopic images for screening non-germinated chili seeds before planting. Experimental results show that GroupCeptionNet achieves 94.66% accuracy and 94.65% F1-score under 1.40 M parameters, outperforming the classical CNN and Transformer models. We also explore the impact of background removal and GroupCeptionNet ablation and variant structures. The visualization results further validate the consistency of the model’s focus areas with human-annotated regions. The proposed dataset and model provide technical references for a low-cost and efficient chili seed screening process.

Introduction

Seed quality is the basis of agricultural productivity and has a direct impact on crop yield and quality. Early studies^1,2 showed that environmental or physical factors easily affect seed quality and decreased yield and quality. Another study³ found that screening seeds for physical and physiological qualities in advance increases crop yield by 15–20%. The Food and Agriculture Organization of the United Nations (FAO) claims high-quality seeds boost crop yields by as much as 100%⁴. Chili pepper (Capsicum spp.) is one of the major cash crops in the world. In 2021, the United States harvested over 31,000 acres of bell peppers and 10,000 acres of chili peppers, totaling more than 536 million dollars⁵. These findings show that improving seed quality is important for increasing crop yields and economic returns.

The germination screening of pepper seeds significantly improved the quality of cultivated seeds. Traditional classification methods such as visual inspection, flotation test and staining methods have limitations⁶. These methods rely on the experience and skills of inspectors, and fatigue makes the results subjective. A study⁷ shows that traditional methods are still effective in some specific cases, but modern agriculture requires greater efficiency and accuracy. Accurate screening of germinated seeds increases land resource utilization rate, yield and cultivation quality, especially for large-scale cultivation⁸. With advanced agricultural production automation, researchers use structured data such as seed length, width and weight or image analysis technology to determine whether chili seeds germinate more effectively⁹. Finer feature extraction necessitates high-quality acquisition equipment and complex algorithms, placing additional demands on the data processing capabilities of the devices.

On the agriculture applications side, the CNN method as a part of agricultural deep learning on seed classification and defect detection is due to its superior skill in feature extraction¹⁰. The traditional CNN model requires a large amount of computing resources and memory, making it difficult to achieve the goals of cheap cost and low power consumption in actual agricultural applications¹¹. Deep learning researchers proposed many lightweight modules and network structure optimization methods¹², which apply them to scenarios that require fast processing or limited resources. These studies optimized the convolution approach and network structure to improve the computational efficiency of the model while retaining the feature extraction ability¹³. The lightweight CNN method provides an effective technical solution for solving practical problems in agriculture¹⁴, which is critical for the advancement of smart agriculture.

Existing chili seed datasets^15–17 are mostly used for species categorization and have not been tagged for germination. The study¹⁸ has shown that seed appearance features are strongly connected with germination ability, and image features are employed as an effective analysis for germination classification. Most of the existing studies use traditional imaging methods for image acquisition, which cannot deeply capture the microscopic details of the seed surface, limiting the accuracy of the model to identify subtle damage and color change¹⁹. Some studies have employed flatbed scanners or hyperspectral devices to take images of chili seeds, but these methods have limited image quality or are more expensive. Macro photography is cheap and simple to take. A smartphone with macro mode or an affordable macro lens module both produce high-quality photographs.

The main contributions of this paper are summarized as follows:

• Dataset Construction: We constructed a standardized macroscopic image dataset for chili seed germination potential classification. This dataset addresses the scarcity of high-resolution micro-feature data in this domain and provides a benchmark for researching fine-grained seed viability.

• Model Innovation: We proposed GroupCeptionNet, a lightweight convolutional neural network designed for resource-constrained agricultural environments. By integrating group convolutions and Inception modules, the model achieves a superior balance between efficiency and performance, attaining 94.66% accuracy with only 1.40 M parameters.

• Comprehensive Validation: We conducted extensive experiments to validate the proposed method, including comparisons with state-of-the-art CNNs and Transformers, ablation studies on image resolution necessity and background removal impact, and interpretability analysis using Grad-CAM to verify the consistency of the model’s focus with human-annotated regions.

The contents of the following sections are organized as follows: Sect.  2 reviews the research on seed classification including various seed data formats analyses and existing classification methods. Section  3 introduces dataset construction, preprocessing and GroupCeptionNet model design. Section  4 describes the experimental setup including dataset partitioning, parameter selection, and experimental arrangement. Section  5 gives experimental results validating the advantages of GroupCeptionNet in classification metrics and resource efficiency. Section  6 provides a detailed discussion of the experimental results. Section  7 concludes the paper and proposes future research directions.

Related work

Seed image germination classification and variety classification both rely on the extraction and analysis of seed surface features. Olaes et al.²⁰ proposed a method based on image processing and fuzzy k-nearest neighbor (KNN)²¹ to realize the classification of pepper seeds. Diosse et al.²² evaluated various classification techniques using the MATLAB toolbox. Weighted KNN and boosted trees have the highest classification accuracy. Loddo et al.²³ designed SeedNet with a simplified network structure, which increased the inference speed while maintaining a classification accuracy of 97.47%. Kurtulmus et al.¹⁵ integrated feature selection with multilayer perceptron (MLP)²⁴. The scanner captures the color, shape and texture features of pepper seeds, resulting in an accuracy of 84.94%. Sabanci et al.¹⁶ combined deep features of ResNet²⁵ with a support vector machine (SVM). This CNN-SVM combination has 99.02% accuracy. The above studies demonstrate the efficiency of the seed image classification method based on surface morphological features. Seed surface characteristics not only identify the variety type but also correlate with germination potential.

The seed germination classification task initially studied transforming image features into structured data for classification. Huang et al.²⁶ developed an automatic sorting system to capture the shape, color and texture features of kale seeds. The positive feature selection increases the final accuracy of the system to 90%. Men et al.²⁷ employed infrared thermal imaging technology to record temperature changes during water absorption. Learning these temperature changes by SVM results in a new lossless way of identifying seed germination. Deep learning models based on CNNs directly learn and recognize image features. For example, Huang et al.²⁸ employed the GoogLeNet²⁹ model to classify corn seed defects with an accuracy of 95%. Zhao et al.³⁰ proposed a real-time recognition system using MobileNetV2³¹. They employed a mobile microscope to collect soybean seed images with an accuracy of 97.84% and an inference speed of 35 FPS, which is suitable for resource-constrained devices. Learning the features of chili seeds using the model provides a solution for germination classification.

The classification of chili seed germination has many similarities with other seeds. Existing research also employs surface morphology or structural features. Tu et al.³² employed machine vision to collect 15 features of chili seeds such as color and shape through a scanner. MLP training delivered a germination classification accuracy of 79.10%. Pornpanomchai et al.³³ developed a pepper seed germination analysis (CSGA) system. The system combines a handheld microscope camera to extract color and texture features. Euclidean distance and the neural network method reach an accuracy of 71.71%. In addition, Hong et al.³⁴ combined hyperspectral and X-ray imaging to evaluate seed germination with a CNN model. The prediction accuracy increased to 92.51% after combining two imaging information. However, the high cost and complexity of such imaging modalities limit their practical application range. In summary, while these studies validate the feasibility of using image analysis for seed germination classification, they reveal several key limitations. Firstly, many methods rely on expensive or complex equipment, hindering their adoption in low-cost, large-scale agricultural settings. Secondly, some models that achieve high accuracy are built on deep, complex architectures that demand significant computational resources, making them unsuitable for deployment on resource-constrained devices. Lastly, there is a scarcity of research focused on creating lightweight models specifically optimized for the fine-grained surface features captured by accessible macro imaging. Therefore, a gap exists for a solution that balances high accuracy with computational efficiency, utilizing easily obtainable high-resolution images.

To address this gap, this paper employs macroscopic image techniques to obtain seed surface features. Macroscopic imaging is more detailed than traditional imaging and less expensive than sophisticated imaging modalities, making it an ideal choice. Building on the exploration of lightweight paths in previous studies, we designed the GroupCeptionNet model. By controlling the convolution depth to reduce redundancy and using a multi-scale approach to maintain feature extraction capabilities, the model’s computational requirements are significantly lowered. This work thus provides an efficient and practical solution for classifying pepper seed germination.

Materials and methods

As shown in Fig. 1, the process of this study is divided into four main steps. Step 1: Images of chili seeds were taken with macro photography to capture subtle surface details. Step 2: Preprocess the collected images. Standardizing image formats to improve model generalizability. Step 3: Extract features from preprocessed images and classify them using the lightweight GroupCeptionNet model. Step 4: Output classification results to evaluate model performance. The following subsections will discuss the specific configurations in the above steps.

Fig. 1.

Overview of the process in this study.

Dataset

The dataset used in this study contains 3840 macro images of the chili pepper seed ‘SuJiaoDaQingXiang F1’, which was purchased from Shuanglong Agricultural Technology Co., Ltd. in Xiao County, China, and stored in a sealed bag. Data collection uses the macro mode of an iPhone 15 Pro with a fixed height of 1.7 cm, a stabilized bracket, and an annular shadowless lamp to ensure uniformity of angle, illumination and high image quality. Each seed was placed on a black PVC matte background board to increase the contrast between the seed and the background. The original images were captured at a resolution of 3024 × 3024 pixels. Through the macro lens, the effective area occupied by a single seed is approximately 1024 × 1024 pixels. This high effective resolution ensures that even after resizing to the model input size (224 × 224), the high-frequency surface details—such as subtle wrinkles and initial seed coat rupture—are well-preserved via downsampling. In contrast, standard photography without a macro lens typically captures the seed at a much lower effective resolution (estimated around 140 × 140 pixels), leading to a loss of these microscopic features. The labels of the dataset were generated based on a standard germination experiment: the seeds were removed from the sealed bag, photographed, and then immersed in water at 30°C for 6 hours. Finally, they were moved to a thermostat at 26°C and placed on a wet paper bed for cultivation. The samples that germinated during the 14-day cultivation period were counted. The target variable for the classification task labels ‘Germinated’ seeds as 1 and ‘non-germinated’ seeds as 0.

Figure 2a shows examples of germinated and non-germinated samples obtained using macro photography. Germinated samples usually have a smooth surface and a complete form. Non-germinated samples have broken surfaces, wrinkles, or local color changes. These surface differences are more pronounced under macro photography. Figure 2b shows the effect of standard photography. The photo distance causes standard photography to capture only the approximate shape and color information of the seeds. These photos make it hard to discriminate between germinated and non-germinated samples solely on fuzzy information. Macro photography provides richer features through clear surface texture and structure. These morphological features assist the model in more precisely identifying seed states of germination.

Fig. 2.

Examples of chili seed samples and comparison of shooting styles.

Table 1 shows the distribution of germinated and non-germinated samples in the pepper seed dataset, distinguishing between ‘untreated’ and ‘treated.’ The dataset includes 2524 untreated seeds, of which 1997 germinated and 527 non-germinated. To balance positive and negative samples, the dataset was augmented with non-germinated samples by intervention, including both physical damage and fungal infection methods. Physical damage in the natural environment is simulated by breaking, perforating and mutilating seeds. The fungal infection method involves placing the seeds with rotten bananas in a damp petri dish. Fusarium and Aspergillus spp. are the most common causes of decay in bananas. These two fungi frequently cause mold and rot in seeds. The intervention method added 569 physically damaged samples and 747 fungal-infected non-germinated seeds for a total of 1316. The total samples in the dataset reached 3840 after the expansion, and the sample distribution was more balanced. A balanced dataset improves the stability and accuracy of the model on germination classification.

Table 1.

Sample distribution of the dataset.

Seed status	Untreated	Treated		Germination total
		Physical damage	Fungal infection
Germinated	1997	–	–	1997
Non-germinated	527	569	747	1843
Treated total	2524	1316		3840

Data preprocessing

Figure 3 illustrates the five data preprocessing steps of this system. These steps effectively improve the training effect and generalization ability of the model. Firstly, the background segmentation method based on deep learning (rembg library)³⁵ removes the background of the original image to reduce unnecessary interference. The contour detection is then utilized to locate the minimum enclosing rectangle of the seed, resulting in a square image containing only the seed body. The cropped images are uniformly resized to 224 × 224 pixels to suit the input criteria of our model and most other comparison models (299 × 299 pixels for InceptionV3³⁶. Next, a random horizontal and vertical flip with a probability of 0.5 and a random ± 10° rotation with a probability of 0.25 increase the diversity of the image. This allows the model to adapt to variations in seed orientation. Color jitters randomly adjust brightness, contrast, and saturation. The adjustment factor was set to 0.25 to represent anticipated variations in lighting conditions. Color jitter keeps the model from becoming overly sensitive to the color features of the seed surface. Pixel values are normalized using the mean and standard deviation of the ImageNet dataset³⁷ before input to the model. A standard distribution of input values improves model training stability.

Fig. 3.

Data preprocessing workflow.

GroupCeptionNet

This paper proposes GroupCeptionNet to increase the efficiency and accuracy of chili seed germination classification. This lightweight model is designed to screen chili seeds for sprouting. As shown in Fig. 1, the core structure of GroupCeptionNet combines the advantages of the ResNeXt block³⁸ and the Inception block²⁹. Each convolution stage contains a core structure (GroupCeption block). Redesigning the depth of the convolution group reduces the parameters while improving the efficiency of the overall model to extract diverse features. The group convolution employed in the ResNeXt block effectively lowers parameters at a fixed depth. Processing group features after input channel segmentation also broaden the feature space. For the standard convolution parameters:

1

Here, and are the height and width of the convolutional kernel, respectively, is the number of channels in the input feature map, and is the number of channels in the output feature map. In group convolution, and are divided into groups, and each group performs a convolution calculation independently, thereby reducing the total parameters to:

2

Group convolutions reduce the parameters to 1/G and allow the model to expand the feature space. The Inception block captures the information from several spatial scales in the seed image through multi-scale feature extraction. The combination of these two modules better adapts to the complexity of chili seed images. Preliminary experimental findings indicate that VGG16³⁹ performs best in terms of accuracy, but it suffers from excessive computational complexity. ResNet18, with the second highest accuracy, has a relatively low parameter count. MobileNetV2³¹ has the fewest parameters but underperforms compared to ResNet18 and VGG16. GroupCeptionNet innovates modules based on ResNet18 to achieve higher accuracy with lower parameters. Table 2 shows the differences in convolutional layer design between GroupCeptionNet and ResNet18.

Table 2.

Comparison of GroupCeptionNet and ResNet18 convolutional layer designs.

Stage	Output	ResNet18	GroupCeptionNet
Conv1	112 × 112	7 × 7, 64, stride 2	7 × 7, 64, stride 2
Conv2	56 × 56	3 × 3, max pool, stride 2	3 × 3, max pool, stride 2
Conv3	28 × 28
Conv4	14 × 14
Conv5	7 × 7
Output	1 × 1	Global average pool 2-d fc, softmax	Global average pool 2-d fc, softmax

Compared with ResNet18, GroupCeptionNet has several optimizations in the convolutional layer design. GroupCeptionNet replaces the first two-layer convolutional block (ResNet block) in each stage with a ResNeXt block. The number of groups is dynamically set to based on the number of channels. This design not only reduces convolution parameters but also solves the problem of whether the convolution channel is too shallow or too deep. The fixed group scheme adopted by ResNeXt has insufficient feature expression due to overly shallow convolution channels in the initial stage. In the final stage, overly deep channels yield high parameters. The fixed depth scheme achieves an appropriate compromise between feature extraction and lightweight. In the improved ResNeXt block, the input features are first extracted by grouping convolution, which is calculated as:

3

Where represents the subset of channels in the input feature that belong to the -th group, each group is convolved separately, and the group outputs are concatenated using feature concatenation to form the final feature. When and do not match, the residual branch uses to adjust the channels to ensure the residual connection operation can be performed effectively. The residual is formulated as:

4

After the first layer of group convolution is calculated, the features are again passed through a second layer of group convolution and combined with the residual branch. The output features of the improved ResNeXt block are calculated as follows:

5

This improved ResNeXt block retains the ability to extract features from multiple layers while considerably reducing the parameters. The second ResNet block in each GroupCeptionNet stage is replaced by an Inception block to improve multi-scale feature capture ability. The optimal channel allocation in the Inception block extracts the information of different spatial scales in proportion. Channel assignment takes into account the role of each path in feature extraction. The allocation scheme similarly achieves a balance between computational efficiency and feature expression. The feature calculation formulas for each subpath are as follows:

6

7

8

9

These features extract different types of information in multi-scale space, and the allocation ratio of output channels (1/4, 1/2, 1/8, 1/8) reflects the fine balance of multi-scale feature extraction. 1/4 of the channels are allocated to for fast downscaling and dimensionality reduction to reduce computational overhead; 1/2 of the channels are allocated to , which gives priority to capturing the most representative mesoscale features in the most significant number of seed images; 1/8 of the channels are allocated to and respectively to ensure the expansion of the receptive field and robust spatial information extraction while avoiding an excessive increase in the overall computational burden. The output features of the Inception block are obtained through a feature stitching operation as follows:

10

The Inception block with optimized channel allocation improves the feature expression ability for chili seed images. The GroupCeptionNet block passes the output of the optimized ResNeXt block to the Inception block. The overall output features of the model are represented as follows:

11

The GroupCeptionNet block design effectively combines the advantages of two types of improved blocks. The GroupCeptionNet model improves the representation of complex image features while keeping the computational overhead low. Each GroupCeptionNet block is considered a stage, following the design of the ResNet18 convolutional stages and iterating through the four stages to extract and enlarge features gradually. The spatial dimensions (height and width) of the feature maps are halved, while the channel depth is doubled after each stage. The feature map starts from 56 × 56 × 64 and finally reaches 7 × 7 × 512 in the fourth stage. The incremental magnification design enables the model to smoothly capture the details and global features.

After feature extraction, the GroupCeptionNet model converts the feature map into a global feature vector via global average pooling. 2-D fully connected layer and a softmax layer combine to produce the final classification. This modular design enables the GroupCeptionNet model to fully adapt to the complexity and diversity of chili seed images.

Experimental settings

In this experiment, the dataset was split into training set, validation set and test set as 6:2:2, and the distribution is shown in Table 3. Notably, the test set was only randomly selected from the images of untreated seeds and 50.00% germinated. This data splitting ensures fairness for each model during the testing phase and more accurately measures the generalization ability in practical applications. The training parameters are set to batch size 32, epochs 40, and learning rate 1e-04. The optimal optimizer (SGD/Adam)^40,41 is selected for each model through multiple experiments and combined with weight decay 5e-05 to prevent overfitting. Optimizers in training use the CrossEntropyLoss function⁴². The final model is saved as the model with the highest F1-score for the entire epoch. Experiments were run on an i7-13700KF CPU, 64GB RAM and an NVIDIA GeForce RTX 4090 GPU with CUDA version 11.8. All models were implemented using PyTorch 2.0.1.

Table 3.

Dataset split and sample distribution.

Set	Germinated	Non-germinated	Germination rate (%)
Training set	1204	1100	52.26
Validation set	409	359	53.26
Test set	384	384	50.00

To rigorously evaluate the model’s performance specifically on detecting natural germination potential, the Test Set was constructed exclusively from ‘Untreated’ (natural) seeds. We randomly selected 384 germinated and 384 non-germinated seeds from the untreated group. Crucially, no artificially intervened samples (physical damage or fungal infection) were included in the Test Set. In contrast, the Training Set incorporated all three categories: untreated seeds, physically damaged seeds, and fungal-infected seeds. This mixed training strategy utilizes the intervened samples as ‘hard negatives’ to help the model learn robust decision boundaries. Consequently, the exclusive use of untreated seeds in the testing phase ensures that the evaluation metrics strictly reflect the model’s capability to identify subtle, naturally occurring non-germination features in real-world scenarios, rather than merely detecting obvious intervention traces.

Furthermore, the decision to use a balanced (50:50) Test Set was necessitated by the extreme class imbalance observed in the natural data (approx. 4:1). A balanced test set prevents the ‘accuracy paradox’ and forces the model to demonstrate equal competence in identifying both germinated and non-germinated seeds, providing a more rigorous assessment than the natural skewed distribution.

This study selected a variety of classic CNNs and Transformer-based models for comparative analysis to comprehensively evaluate the GroupCeptionNet model’s lightweight and classification performance. The comparison models include AlexNet⁴³, VGG series (11/13/16/19)³⁹, ResNet series (18/34/50/101/152)²⁵, ResNeXt18³⁸, MobileNet series (V2/V4_Small/V4_Medium/V4_Large)^31,44, GoogLeNet²⁹, InceptionV3³⁶, as well as VisionTransformer⁴⁵ and SwinTransformerV2⁴⁶. The VGG series has a simple structure, but the last two fully connected layers make the parameters and memory footprint large. The ResNet series achieves a relatively low computational cost while maintaining a deeper structure by introducing residual links. The MobileNet series optimizes the convolutional structure for embedded scenarios, significantly reducing the parameters and computational complexity. The Transformer series uses self-attention mechanisms to capture long-distance feature dependencies. Compared with other comparison models, GroupCeptionNet achieves the smallest parameters, reducing resource consumption.

This study sets experiments of background removal and retention to analyze to better understand the influence of retaining backgrounds on model accuracy. Background information interferes with the extraction of chili seed features. The frosted background of the images used in this study also contains a certain amount of noise. The experiments compare the model performance under different image background preprocessing conditions. A comparison between background removal and retention experiments determines its necessity in data preprocessing.

Experiments also include a series of ablation and variant experiments for GroupCeption blocks. This experiment analyzes the specific effects of different blocks or structures on performance and parameters. Table 4 shows the parameters and structural descriptions of four ablations and four variant structures. By comparing and analyzing these ablation and variant models, the contribution of each block in the lightweight design and feature extraction of the GroupCeptionNet model is further evaluated.

Table 4.

Ablation and variant experiment configurations for GroupCeptionNet.

	Model	Parameters (M)	Structure description
Ablation	Incep-no_Res	0.59	One inception block only
	Res-no_Incep	4.90	One ResNet block only
	Group-no_Incep	0.74	One ResNeXt block only
	Group-Incep (1,3,3, MP)	1.40	ResNeXt block with Inception (3 × 3 instead of 5 × 5 kernel) block
Variant	Incep-Res	6.86	Inception block followed by ResNet block
	Res-Incep	5.61	ResNet block followed by Inception block
	Incep-Group	1.15	Inception block followed by ResNeXt block
	Group-Incep (GroupCeptionNet)	1.44	ResNeXt block with Inception block

This study complements the confusion matrix analysis and the heat map visualization experiment with Intersection over Union (IoU) to comprehensively analyze the performance of the GroupCeptionNet model on different dimensions. Confusion matrix analysis shows the classification accuracy and balance of the model among categories. This analysis provides insight into the bias and misclassification trends during the classification task. Heatmap visualization and IoU evaluation quantify the degree of overlap between the area of attention and annotation of the model. The areas with a heatmap value higher than 0.7 are considered attention areas.

In this study, accuracy, precision, recall and F1-score are key indicators for assessing model performance. Accuracy measures the overall classification accuracy of the model for all samples. Precision indicates the reliability of the model in predicting positive samples, and recall indicates the proportion of all actual positive samples that the model correctly identifies. The F1-score reflects the balance of the model in identifying positive and negative samples in a classification task. These indicators provide an essential basis for analyzing model performance.

Results

Model performance comparison

Table 5 shows the classification results and performance requirements on the test set, corresponding to the actual application of each model in the classification of the chili seed germination task. The results include accuracy (Acc), precision (Prec), recall (Rec) and F1-score (Fs), as well as memory usage and GFLOP. Figure 4 visually shows the relationship between GFLOPs and accuracy for each model.

Table 5.

Comparison of the performance of various models on the chili seed germination dataset. Bold numerical values of the model performance

Model	Acc	Prec	Rec	Fs	Memory usage (MB)	GFLOPs
VisionTransformer	88.28	88.94	88.28	88.23	739.12	16.88
MobileNetV4_Small	90.23	91.42	90.23	90.16	93.72	0.18
ResNet152	91.54	92.54	91.54	91.49	828.97	11.60
MobileNetV4_Medium	91.54	92.06	91.54	91.51	286.88	0.83
MobileNetV4_Large	91.54	91.97	91.54	91.51	531.77	2.17
GoogLeNet	91.93	92.77	91.93	91.89	116.04	1.51
ResNet101	92.06	92.70	92.06	92.03	592.44	7.86
MobileNetV2	92.32	92.81	92.32	92.30	161.93	0.33
ResNet34	92.45	93.12	92.45	92.42	178.05	3.68
ResNet50	92.45	92.91	92.45	92.43	376.82	4.13
SwinTransformerV2	92.45	92.59	92.45	92.44	343.11	3.34
AlexNet	92.58	93.23	92.58	92.55	178.05	0.71
ResNeXt18	92.84	93.33	92.84	92.82	287.08	2.51
InceptionV3	93.23	93.61	93.23	93.21	211.68	2.86
VGG19	93.49	94.07	93.49	93.47	771.70	19.63
ResNet18	93.62	94.12	93.62	93.60	106.00	1.82
VGG16	93.62	94.12	93.62	93.60	731.54	15.47
VGG11	94.14	94.55	94.14	94.13	617.18	7.61
VGG13	94.40	94.82	94.40	94.39	691.38	11.30
GroupCeptionNet	94.66	94.75	94.66	94.65	65.65	0.47

Fig. 4.

Relationship between GFLOPs and accuracy.

Table 5 and Fig. 4 show that the GroupCeptionNet model achieved the highest classification performance on several key evaluation metrics, with accuracy and F1-score of 94.66% and 94.65%, respectively, demonstrating a significant advantage in the classifying chili seed germination task. It also demonstrates excellent lightweight and efficiency with 1.4 M parameters, 65.65 MB memory usage and 0.47 GFLOPs. This result shows that GroupCeptionNet not only has efficient feature extraction capabilities but is also highly applicable in resource-constrained environments.

In the VGG series, VGG13 and VGG16 achieved accuracies of 94.40% and 93.62%, respectively, proving that a simple structure is more effective for this task. The high parameter counts of 128.96 M for VGG13 and 134.27 M for VGG16 and the large memory footprint of 691.38 MB and 731.54 MB limit their application flexibility in resource-constrained environments.

The ResNet series offers a wide range of options in terms of performance and lightweight. Among them, ResNet18 and ResNeXt18, which have the advantage of being lightweight, have accuracies of 93.62% and 92.84%, respectively, and GFLOPs of 1.82 and 2.51. This result proves that residual connection convolutional layers enhance the feature learning ability of CNN models. Grouped convolutions optimize computational efficiency, making them suitable for broader applications.

The MobileNet series of models achieves better efficiency through depthwise separable convolution. MobileNetV4_Medium and MobileNetV2 have accuracies of 91.54% and 92.32%, respectively, which are slightly inferior to GroupCeptionNet. MobileNet has high application potential in resource-limited scenarios. In particular, MobileNetV4_Small has 0.18 GFLOPs, much lower than other models.

GoogLeNet and InceptionV3 rely on multi-scale feature extraction to achieve accuracies of 91.93% and 93.23%, respectively, demonstrating the effectiveness of multi-scale feature capture. Compared to GroupCeptionNet and VGG13, the more complex convolutional stage design did not significantly improve performance. AlexNet, CNN’s foundation, had a 92.58% accuracy in this task. These three models again demonstrate that relatively simple structures are more effective in this task.

The GroupCeptionNet model achieves a better balance between accuracy, lightweight and resource efficiency than other models. This result is attributed to the combination of ResNeXt block and Inception block within the stage of ResNet18. GroupCeptionNet meets the high-performance requirements of classifying chili seed germination tasks and provides a reliable technical reference for this application scenario.

Background removal experiment results

Table 6 compares the classification performance of different models under background removal and background retention preprocessing conditions to analyze the role of background removal in classifying chili seed germination tasks. Background removal improves the model’s classification performance, especially in accuracy and F1-score, which are, on average, 1.71% and 1.73% higher. This result shows that removing background information in this task helps the model focus more on the characteristics of chili seeds.

Table 6.

Background removal vs. retention.

Model	Background removal		Background retention		Improvement
	Acc	Fs	Acc	Fs	Acc	Fs
VisionTransformer	88.28	88.23	85.94	85.82	2.34	2.41
MobileNetV4_Small	90.23	90.16	88.67	88.56	1.56	1.60
ResNet152	91.54	91.49	90.62	90.56	0.92	0.93
MobileNetV4_Medium	91.54	91.51	91.41	91.40	0.13	0.11
MobileNetV4_Large	91.54	91.51	84.38	84.30	7.16	7.21
GoogLeNet	91.93	91.89	90.23	90.17	1.70	1.72
ResNet101	92.06	92.03	91.93	91.89	0.13	0.14
MobileNetV2	92.32	92.30	91.67	91.62	0.65	0.68
ResNet34	92.45	92.42	92.32	92.29	0.13	0.13
ResNet50	92.45	92.43	92.32	92.29	0.13	0.14
SwinTransformerV2	92.45	92.44	87.37	87.32	5.08	5.12
AlexNet	92.58	92.55	91.93	91.93	0.65	0.62
ResNeXt18	92.84	92.82	90.76	90.72	2.08	2.10
InceptionV3	93.23	93.21	92.32	92.29	0.91	0.92
VGG19	93.49	93.47	92.19	92.15	1.30	1.32
ResNet18	93.62	93.60	88.93	88.81	4.69	4.79
VGG16	93.62	93.60	92.58	92.55	1.04	1.05
VGG11	94.14	94.13	93.10	93.08	1.04	1.05
VGG13	94.40	94.39	93.62	93.61	0.78	0.78
GroupCeptionNet	94.66	94.65	92.97	92.96	1.69	1.69
Average improvement					1.71	1.73

In this study, the model classifies by learning the delicate features of the seed surface, and the background information obscures these fine features to some extent. After applying background removal, GroupCeptionNet’s accuracy and F1-score reached 94.66% and 94.65%, respectively, an improvement of 1.69% and 1.69% compared to when the background was retained. Similarly, classical CNN models such as VGG, ResNet, and GoogLeNet also achieved performance improvements of varying degrees after background removal. For example, VGG19’s accuracy and F1-score improved by 1.30% and 1.32%, respectively, while ResNeXt18 achieved 2.08% and 2.10% improvements.

Although all models have gained better performance due to background removal, the degree of performance improvement varies due to differences in model architecture. For example, the accuracy and F1-score of SwinTransformerV2, which uses a multi-head self-attention mechanism, improved by 5.08% and 5.12%, respectively, while AlexNet’s improvement was relatively small, at only 0.65%. These results show that background removal significantly affects models primarily using deep convolutions and self-attention mechanisms. This result shows that removing the background during data preprocessing can improve the classification accuracy of the model.

In addition to accuracy improvements, we evaluated the computational efficiency of the background removal step. On our experimental setup, the model inference for background removal takes an average of 0.0761 s per image. This rapid processing speed indicates that the background removal step introduces negligible latency, making it highly suitable for real-time deployment in automated agricultural sorting systems.

Ablation and variant experiment results for GroupCeptionNet

Table 7 shows the results of the ablation and variant schemes of the GroupCeptionNet model. The experiments include four ablations and four variant experiments to evaluate the impact of ResNet blocks, Inception blocks, and ResNeXt blocks on classification performance. These results are used to analyze the specific role of each module role in model lightweight and classification performance.

Table 7.

Ablation and variant experiment results for GroupCeptionNet.

	Model	Memory usage (MB)	Acc
Ablation	Incep-no_Res	148.41	91.02
	Res-no_Incep	61.98	89.71
	Group-no_Incep	46.09	90.89
	Group-Incep (1,3,3, MP)	65.65	94.14
Variant	Incep-Res	333.11	92.97
	Res-Incep	81.70	93.75
	Incep-Group	311.31	94.01
	Group-Incep (GroupCeptionNet)	65.81	94.66

In the ablation experiment, Incep-no_Res and Res-no_Incep retain the Inception and ResNet blocks, respectively, with parameters of 0.34 M and 4.91 M, and memory usage of 148.41 MB and 61.98 MB, but the accuracy is significantly reduced, at 91.02% and 89.71%, respectively. Similarly, the parameters for Group-no_Incep is 0.74 M. The memory usage is further reduced to 46.09 MB, but the accuracy is still only 90.89%. These results suggest that either block alone struggles to achieve the desired classification performance. Group-Incep (1,3,3,MP) removes the 5 × 5 convolutional kernel, reducing parameters by 0.04 M and memory usage by 0. 24 MB, with a slight decrease in accuracy to 94.14%. This ablation suggests that the 5 × 5 convolutional kernel increases the extraction of fine features on the seed surface but produces a smaller parameter and memory overhead. However, since only 1/8 is allocated to the 5 × 5 convolution channel, the resulting parameter and memory overheads are relatively limited.

In the variant experiment, Res-Incep reduces memory usage from 333.11 MB to 81.70 MB compared to Incep-Res, and parameters are reduced from 6.61 M to 5.43 M, while the accuracy increases from 92.97% to 93.75%. Similarly, Group-Incep reduces memory usage from 311.31 MB to 65.81 MB compared to Incep-Group, and the accuracy increases from 94.01% to 94.66%. This trend shows that placing the Inception block after the ResNet or ResNeXt block significantly reduces the memory usage and parameters of the model while ensuring efficient feature extraction capabilities. The resource efficiency of the model is further optimized by replacing the standard convolution with a group convolution. Compared with Res-Incep, Group-Incep not only reduces memory usage from 81.70 MB to 65.81 MB but also reduces parameters from 5.43 M to 1.44 M, while the accuracy is improved to 94.66%. This result further proves that group convolution effectively improves feature extraction capability by expanding the feature space while significantly reducing the resource requirements of the model.

Based on the above results, the GroupCeption block finally chooses the Group-Incep structure to achieve the highest classification performance with lower memory overhead and parameters. The GroupCeption block combination design was optimized and demonstrated through variant and ablation experiments to achieve the ideal balance between classification performance and resource efficiency, providing an efficient and reliable model structure for classifying chili seed germination.

Analysis of image resolution necessity

To empirically verify the necessity of high-resolution macro images for detecting fine seed germination features, we conducted an ablation study focusing on the effective acquisition resolution.

As described in Sect.  3.1, our macro imaging setup captures an effective seed area of approximately 1024 × 1024 pixels. To simulate non-macro acquisition scenarios, we employed a “Simulation of Acquisition Resolution” strategy during the validation phase. We downsampled the cropped seed images to effective resolutions of 140 × 140 pixels (simulating standard smartphone photography) and 70 × 70 pixels (simulating low-quality or distant capture) before resizing them to the model’s input size of 224 × 224 pixels. This process mimics the physical loss of detail inherent in non-macro photography while maintaining the model architecture constant.

The results in Table 8 demonstrate a clear dependency on image resolution. When the effective resolution drops to 140 pixels (standard photography), the accuracy decreases by 2.52% to 92.14%, falling to a level comparable with conventional CNNs lacking macro-optimized features. A further reduction to 70 pixels leads to a sharp performance degradation (Acc 87.89%).

Table 8.

Performance comparison under different simulated acquisition resolutions.

Simulated scenario	Effective seed resolution	Accuracy (%)	F1-score (%)
Macro photography (ours)	~ 1024 px (Resized to 224)	94.66	94.65
Standard photography	140 px (Upscaled to 224)	92.14	92.08
Low-quality/far distance	70 px (Upscaled to 224)	87.89	87.76

This ablation study validates that GroupCeptionNet explicitly utilizes the fine epidermal structural features, such as micro-cracks and subtle texture alterations, which are solely retained in high - resolution macro images. While the global shape information available in lower resolutions supports basic classification, the superior accuracy of our model is directly attributed to the high-frequency details provided by macro photography.

Confusion matrix analysis

Figure 5 shows the confusion matrices for the top four model series regarding classification accuracy, GroupCeptionNet, VGG13, ResNet18 and InceptionV3, on classifying the germination of chili seeds. These confusion matrices reveal the specific performance of each model in classifying ‘germinated’ and ‘non-germinated’ samples, especially the difference in classification balance. VGG13, ResNet18, and InceptionV3 exhibit a tendency to misclassify ‘non-germinated’ samples as ‘germinated’ while misclassifying ‘germinated’ samples less frequently. The misclassification rates of VGG13 and ResNet18 were 10.4% and 11.7% for ‘non-germinated’ samples and only 0.8% and 1.0% for ‘germinated’ samples. This phenomenon suggests that these models tend to conservatively predict the ‘germinated’ category in the classification task.

Fig. 5.

Confusion matrix comparison.

In contrast, GroupCeptionNet achieves a better balance between the ‘germinated’ and ‘non-germinated’ categories, with a 7.6% misclassification rate for the ‘germinated’ category and a 3.1% misclassification rate for the ‘non-germinated’ category. This result shows that GroupCeptionNet’s classification accuracy is balanced across different categories, with no extreme bias towards one category. These confusion matrices demonstrate the model’s ability to differentiate between samples, confirming GroupCeptionNet’s strength in classification stability.

Heatmap visualization and IoU evaluation

To interpret the models’ decision-making process and localize discriminative regions, we utilized the Gradient-weighted Class Activation Mapping (Grad-CAM) technique. Specifically, for all compared models, we selected the final convolutional layer of the last feature extraction stage as the target layer, as it preserves the most abstract semantic spatial information before the global average pooling or fully connected layers. The class activation maps were generated by computing the gradients of the target class score with respect to the feature maps, applying a ReLU activation to suppress negative contributions, and upsampling the resulting heatmap to the original input resolution for visualization.

Figure 6 shows the heatmap visualization comparison of four representative seed samples and quantitatively evaluates the model by calculating the Intersection over Union (IoU) of the attention area and annotated area. The red box indicates the annotated area, and the blue boxes indicate the attention areas of the model (threshold > 0.7).

Fig. 6.

Visualization of final feature layer heatmap with annotated area.

To quantitatively evaluate the model’s feature localization ability, we defined the ground truth ‘annotated regions’ based on the semantic characteristics of the seeds:

• For germinated seeds: The entire seed body is annotated as the region of interest. Since these seeds lack specific defects, their morphological completeness and healthy surface texture are the key semantic features for classification.

• For non-germinated seeds: Only the specific defective regions (e.g., mold spots, rot areas, cracks, or discoloration) are annotated. These local anomalies are the discriminative features that indicate non-viability.

Based on this protocol, we calculated the IoU between the model’s heatmap attention area and these annotated regions.

Germ.1 has a smooth and uniform color overlay on the seed surface, and the annotated area covers the entire seed. GroupCeptionNet exhibits a broad area of attention, focusing on the center of the seed and covering the annotated area well. VGG13 has a very restricted area of attention, focusing only on the upper edge of the seed. ResNet18 focuses on the right side, while InceptionV3 has an attention area similar to GroupCeptionNet but slightly shifted.

The surface of the Non-G.1 seed bears a large range of small dark spots, and the annotated area covers these spots. The attention areas of GroupCeptionNet, ResNet18, and InceptionV3 are all concentrated in the center of the seed; however, ResNet18 is biased towards the edge, extending beyond the boundary. VGG13 covers only a small portion of the upper edge without overlapping with the annotated area.

Non-G.2 seed has a small concentrated area of decay on the surface. VGG13 excels on this type of seed, with an area of focus closely covering the decayed area, achieving the highest IoU. GroupCeptionNet has a larger focus area that covers the annotated region, resulting in the second-highest IoU. ResNet18 and InceptionV3 show some offset in the focus region, resulting in lower IoU scores.

Non-G.3 seed is partially broken in the center; the surface shows significant porosity due to drying, and there is a small area of discoloration in the lower right corner. GroupCeptionNet maintains global attention, covering the central broken area while also attending to the discolored area. ResNet18 pays attention to both regions but focuses mainly on the discolored part, with some attention spilling beyond the boundary. VGG13 focuses strictly on the discolored area. InceptionV3 focuses on the middle of the broken and discolored areas.

From the quantitative IoU evaluation, GroupCeptionNet demonstrates a high degree of consistency with human-annotated regions, maintaining an effective focus regardless of whether the key features are broadly distributed or locally concentrated. VGG13 consistently shrinks the attention area, an approach that is effective for specific concentrated defects but leads to instability in global feature capture. ResNet18 and InceptionV3 exhibit relatively scattered or shifted attention areas, resulting in overall lower IoU scores compared to GroupCeptionNet.

Discussion

Experimental results show that GroupCeptionNet’s design performs well in classifying pepper seed germination. The combination of the improved ResNeXt block with the Inception block is the key to performance improvement. Grouped convolution significantly reduces the parameters and the amount of computation. By limiting the channel sense field range of each convolutional kernel, the feature independence of the model is improved, and redundancy between features is reduced. GroupCeptionNet dynamically adjusts the number of groups. The shallow stage focuses on basic feature extraction, and the deep stage extends the diversity of the feature space. This design achieves an accurate capture of pepper seed features.

The multi-scale feature aggregation of Inception blocks further enhances the model’s ability to capture complex features. Compared to the complex parameters designed by GoogLeNet for the ImageNet dataset with 1 K categories, GroupCeptionNet reduces redundancy in the two-category pepper seed germination classification task. The GroupCeptionNet design prioritizes mesoscale features as the primary target, treats large-scale features as secondary, and utilizes small-scale features for dimensionality reduction. Such a channel allocation strategy avoids unnecessary computational overhead while preserving multi-scale information.

The ablation and variant experiments together reveal the need for synergy between grouped convolutional blocks and Inception blocks. The grouped convolutional block provides diversified feature representations by reducing parameter redundancy and enhancing feature independence, while the Inception block strengthens the model’s ability to capture complex features through the aggregation of multi-scale features. The combination of the two blocks not only overcomes the performance limitations of a single block but also realizes the feature complementarity between the blocks.

GroupCeptionNet’s design demonstrates a high degree of adaptability in classifying pepper seed germination. This lightweight strategy does not come at the expense of classification performance, but by dynamically adjusting the number of subgroups and optimizing the feature extraction module, it accurately captures the key features on the surface of chili seeds while avoiding unnecessary computational redundancy. The analysis of the visualization results further validates the advantages of GroupCeptionNet in locating critical regions and classification balance. This lightweight design demonstrates GroupCeptionNet’s potential to operate efficiently in real-world scenarios like resource-constrained or embedded devices.

In summary, the GroupCeptionNet design fully demonstrates the effectiveness of grouped convolution and multi-scale feature extraction in classifying chili seed germination. Through the innovative combination of ResNeXt and Inception blocks, GroupCeptionNet not only realizes lightweight and efficient feature expression but also demonstrates deep adaptability to the task requirements. Both in terms of classification performance in experimental results and the consistency of focus areas with human-annotated regions in visualization results, GroupCeptionNet demonstrates its significant advantages as a model customized for binary classification tasks. These features make it an efficient and reliable solution for seed germination classification and provide a solid foundation for further developments in this field.

Conclusion

In this paper, a lightweight GroupCeptionNet model is proposed and successfully applied to classify a macroscopic image dataset of pepper seed germination. The model achieves an ideal balance between resource efficiency and classification performance by innovatively combining ResNeXt and Inception blocks. Experimental results show that GroupCeptionNet outperforms multiple comparative models in terms of accuracy and classification balance, while keeping the number of references and memory requirements low, demonstrating excellent adaptability and reliability. The study proves that a reasonably designed combination of modules and an efficient preprocessing process significantly improves the performance of the seed classification task. This provides an efficient and reliable solution for seed quality inspection in smart agriculture, especially in application scenarios such as resource-constrained or embedded devices.

Future research can further optimize the model structure and explore more ways of combining modules to enhance classification performance and model adaptability. Meanwhile, expanding the diversity of datasets, especially for the collection of unprocessed negative samples, will help enhance the generalization ability of the model. In addition, applying the process proposed in this paper to automated scenarios in real production and developing portable systems for embedded devices may provide broader support and application prospects for agricultural seed quality testing.

Author contributions

Conceptualization, C.A. and T.X.; methodology, C.A.; software, C.A.; validation, C.A. and T.X.; formal analysis, C.A.; investigation, C.A.; resources, T.X.; data curation, C.A.; writing---original draft preparation, C.A.; writing---review and editing, T.X.; visualization, C.A.; supervision, T.X.; project administration, T.X.; funding acquisition, T.X. All authors have read and agreed to the published version of the manuscript.

Funding

None.

Data availability

The data presented in this study have been deposited in the FigShare repository at 10.6084/m9.figshare.30406561. The dataset is currently under embargo (private) but is available for peer review at the following private link: [https://figshare.com/s/85de265837f40eb28cf7](https:/figshare.com/s/85de265837f40eb28cf7) . The data will be made publicly available upon acceptance of the manuscript.The source code for the proposed GroupCeptionNet model, including data preprocessing and training scripts, is openly available on GitHub at: [https://github.com/418109764/GroupCeptionNet](https:/github.com/418109764/GroupCeptionNet).

Declarations

Competing interests

The authors declare no competing interests.