Research on an industrial liquid level detection method based on an improved SSD

Abstract

To resolve the challenge of balancing computational efficiency and accuracy in liquid level detection models within industrial environments, this study proposes a lightweight SSD algorithm tailored for edge devices. The proposed algorithm facilitates real-time liquid level monitoring and anomaly early warning, thereby mitigating the risk of production incidents. The algorithm first builds a lightweight feature extraction network based on MobileNetV2 and substantially reduces model complexity and computational energy consumption. We propose a Multi-Scale Dilated Inverted Residual Block (MS-DIRB) to mitigate fine-grained information loss during downsampling. By integrating multi-dilated convolutions and a hierarchical feature reuse strategy, the module enhances multi-scale representational capacity while maintaining computational efficiency. Furthermore, a Triplet Attention mechanism is embedded into critical bottleneck structures, establishing a three-dimensional dynamic feature calibration framework that improves discriminative feature learning while maintaining model compactness. Experimental results demonstrate that the proposed algorithm achieves 96.65% mAP, representing a 3.57 percentage-point improvement over the baseline SSD architecture, while reducing model parameters by 81.58%. These experimental outcomes demonstrate the effectiveness of the improved algorithm and its suitability as a real-time detection model in industrial edge computing applications.

Introduction

With the growing emphasis on energy conservation, emission reduction, and green manufacturing, optimizing liquid level detection technology has become a critical means of reducing production costs and improving energy efficiency. In practical applications such as agricultural irrigation, chemical reactors, and energy storage tanks, traditional detection methods often suffer from limited environmental adaptability and high maintenance requirements, resulting in inadequate real-time monitoring and an increased risk of safety incidents. Reference^1–3 adopted traditional computer vision-based liquid level detection techniques, relying on handcrafted features and algorithms for image processing and analysis. These methods suffer from poor adaptability to environmental variations and difficulties in handling complex tasks, limiting their broader applicability. In contrast, detection techniques driven by deep learning employ convolutional neural networks (CNNs) to learn features automatically, offering strong generalization capabilities and end-to-end training advantages. These advancements address the shortcomings of traditional approaches and significantly enhance the performance of computer vision systems, particularly in complex and dynamic environments where deep learning demonstrates unparalleled superiority^4,5. Consequently, intelligent liquid level detection technologies based on deep learning have gradually become mainstream. However, existing models exhibit excessive parameter sizes, making them challenging to deploy on edge devices^6,7. Moreover, liquid level detection in industrial settings faces complex and volatile environments (e.g., varying illumination, diverse container materials, liquid reflections). In addressing these challenges, the research focus has shifted from merely pursuing accuracy to enhancing the model’s robustness and adaptability in such complex scenarios. In pipeline defect diagnosis, for instance, Lei Wang et al.^8,9 introduced a “mechanism-aided bilevel knowledge transfer” framework and a subsequent “biunbiased neural network” method, respectively. This series of work integrates physical mechanisms with unbiased estimation, thereby providing a systematic solution to overcome data domain shift and improve model generalization in unknown scenarios. Xiangkai Shen et al.¹⁰ demonstrated the significant potential of multi-task collaborative optimization and physics-informed embedding for enhancing the robustness of model decisions through their “task-oriented physical collaborative network” (TOPC-Net). The core concepts of these studies have collectively informed the design and provided a valuable theoretical framework for this research to address the challenges of liquid level recognition in complex environments. Consequently, designing an object detection model that is lightweight, high-precision, and highly robust is paramount for achieving reliable real-time liquid level monitoring.

Object detection algorithms can be broadly classified into two-stage and one-stage approaches. The main workflow of two-stage detectors consists of candidate region generation¹¹ and subsequent classification and localization of these regions; representative networks include R-CNN¹², Fast R-CNN¹³ and Faster R-CNN¹⁴. One-stage detectors differ by predicting object categories and regressing bounding boxes in a single step over the full image, using a unified pass that yields faster inference and superior performance in real-time applications. Prominent one-stage models include SSD^15–17, YOLO^18,19, and RetinaNet²⁰. Jianfeng Liang et al.²¹ proposed a liquid level detection method based on Ghost-YOLOv4 . By replacing the CSPDarknet53 backbone with the Ghost backbone and co-optimizing the regression loss function and data enhancement strategy, their method reduces the model parameters while improving the detection accuracy. J. A. Syed et al.²² achieved high accuracy recognition of transparent containers and liquids through training a YOLOv8 model. While these methods are effective for transparent containers, they may face challenges when applied to other container types or complex environments. Guangchao Qiao et al.²³ introduced a collaborative YOLOv5s and image processing-based water level measurement method, which extracts gauge regions and scale characters via deep learning to calculate water level elevation, demonstrating strong robustness. While the method performs effectively in wastewater and sewage environments, its ability to detect calibration marks in clear liquids may be compromised by background visual interference. This limitation suggests reduced adaptability when applied across diverse fluid types.

To address the aforementioned issue, we propose a lightweight object detection model that combines strong applicability and high accuracy^24–27. In this paper, SSD is used as a framework, and MobileNetV2²⁸ is used to replace the VGG16²⁹ network, which significantly decreases the parameter count; the Multi-Scale Dilated Inverted Residual Block (MS-DIRB) is designed for downsampling, which strengthens the model’s ability to capture features across multiple spatial scales. Finally, the Triplet Attention mechanism is selectively added to part of the backbone network to capture cross-dimensional information through channel-space-channel three-way interaction, which dynamically calibrates the feature weights and enhances the feature expression capability, thus contributing to the model robustness and accuracy.

Background and related work

The classical SSD model constructs a multi-scale detection framework based on the VGG16 backbone network. By employing a hierarchical detection strategy with predefined anchor boxes, multi-scale aspect ratios are designed to cover objects of various sizes and non-maximum suppression (NMS) is utilized to filter the prediction box^30,31. This single-stage approach circumvents the complex region proposal generation and post-processing steps inherent in two-stage methods, thereby improving target matching efficiency. Although the classical SSD model achieves satisfactory detection performance, it faces issues including an excessive parameter count, substantial computational overhead, and slow inference speed^32–34, which render it incompatible with the real-time constraints and stringent storage limitations characteristic of edge deployment scenarios. The classical single-shot multi-box detector (SSD) network is illustrated in Fig. 1.

Fig. 1.

Classical SSD object detection model network architecture.

Improved SSD object detection network

Lightweighting of backbone networks

To mitigate the high computational cost of the original SSD framework (which employs VGG16 as its backbone), we replace the backbone with MobileNetV2, a lightweight architecture optimized^35,36 for efficiency through depthwise separable convolutions. Depthwise separable convolution decomposes the standard convolution into a cascade operation of depthwise convolution and pointwise convolution, realizes the compression of computation through the parameter decoupling strategy, and provides a lightweight solution for low-power deployment in edge computing scenarios. The depthwise separable convolution is illustrated in Fig. 2. The MobileNetV2 network introduces the inverse residual module to alleviate the feature degradation problem^37–39. Its core structure follows the three-stage design of “expansion-feature extraction-compression”, which retains the ability to express high-dimensional feature space while compressing computational complexity. The module only has a shortcut connection when stride = 1 and the input and output feature matrices are of the same shape, and downsampling is performed directly when stride = 2, as demonstrated in Fig. 3. Furthermore, replacing the original ReLU activation function with ReLU6 in the inverted residual module helps to mitigate gradient explosion and improve training stability.

Fig. 2.

Depthwise separable convolution.

Fig. 3.

Inverted residual structure of MobileNetV2.

The improved SSD architecture, built upon the MobileNetV2 backbone, integrates multi-scale feature fusion across bottleneck5_3, conv18, and four extra feature layers. This design enables robust multi-scale object detection while maintaining a lightweight structure, effectively balancing model precision and computational demand. Especially appropriate for platforms with limited hardware capacity (e.g., embedded systems and mobile devices), the complete architecture is illustrated in Fig. 4.

Fig. 4.

Network architecture of the MobileNetV2-SSD.

Proposal of the Multi-Scale Dilated Inverted Residual Block

Since the MobileNetV2 inverted residual block employing stride 2 lacks shortcut connections, which resulting in the loss of downsampling information and degradation of shallow gradient flow, this paper proposes the Multi-Scale Dilated Inverted Residual block. Building upon the stride-2 inverted residual block,the design integrates the idea of Atrous Spatial Pyramid Pooling (ASPP)^40–42 and leverages the lightweight nature of depthwise separable convolutions to reduce parameter complexity. Specifically, we embed parallel dilated convolution groups (with dilation rates d = [1,2,3]) to design multi-dilated depthwise separable convolutions. This architecture facilitates multi-scale feature extraction while expanding the receptive field to incorporate richer contextual information, thereby addressing the inherent limitations of spatial detail loss in conventional downsampling operations. The dilated depthwise separable convolution with d = 2 is illustrated in Fig. 5.

Fig. 5.

Dilated depthwise separable convolution.

The designed MS-DIRB comprises a multi-branch dilated depthwise separable convolutional layer and adaptive shortcut connections for dimensional adjustment. After performing cross-receptive field feature extraction via three parallel dilated depthwise separable convolutions, the framework employs channel concatenation followed by 1 1 convolution to achieve cross-channel mapping and dimensionality compression. A shortcut connection has been added to address network degradation and gradient dispersion. Due to input–output dimension mismatch, a shortcut branch with 1 1 convolution performs spatial downsampling and channel alignment. Residual addition (Add) achieves cross-scale feature fusion, enhancing shallow gradient response efficiency and suppressing feature loss through projection mapping. Figure 6 illustrates the Multi-Scale Dilated Inverted Residual Block.

Fig. 6.

Multi-Scale Dilated Inverted Residual Block with a stride of 2.

Citation of the Triplet Attention mechanism

To address the problem of the insufficient feature characterization ability of depthwise separable convolution in MobileNetV2 due to limited cross-channel information interaction, this paper introduces the Triplet Attention lightweight attention mechanism⁴³. This mechanism establishes global dependencies through cross-dimensional three-way interaction, enabling full integration of spatial and channel information to enhance attention weights in key regions and to realize an end-to-end adaptive feature enhancement system^44–46. Triplet Attention consists of three parallel branches as shown in Fig. 7. In branch 1, the height and channel interaction branch, the input tensor C H W (number of channels C, height H, and width W) is rotated counterclockwise by 90 along the H-axis, then processed by Z-Pool and standard convolutional layers. After applying the sigmoid activation to generate attention weights, an inverse rotation is performed to restore the original dimensions. Branch 2 is an interactive fusion between the channel and width dimensions, which is computed in the same way as the first branch. Branch 3 establishes the interaction between the height and width dimensions by directly performing Z-pooling and convolution operations on the original tensor. The resulting features are passed through a sigmoid activation to compute attention weights. Finally, the tensors produced by the three branches are aggregated to form the output tensor Y.

Fig. 7.

Schematic diagram of Triplet Attention.

In the computation process, the Z-pool layer conducts max-pooling and average-pooling operations on the input tensor, then concatenates the two pooled tensors. This reduces feature dimensionality and computational effort, thereby accelerating the training and inference phases of the model. It is denoted as equation (1):

1

and correspond to the maximum and average pooling operations, respectively.

Equation (2) defines the output tensor Y produced by the last three branches.

2

The symbol denotes the sigmoid activation function, while , and correspond to convolution operations. The tensors and are generated by rotating the input tensor degrees counterclockwise along the vertical (H) and horizontal (W) axes, respectively. is derived through a direct Z-pooling operation. An overline indicates that the transformed tensor undergoes a reverse rotation–clockwise along the same axis used previously.

Triplet Attention constructs a joint spatial-channel attention mechanism through three-dimensional cross-dimensional interactions, effectively mitigating limitations of MobileNetV2 in feature representation⁴⁷. Considering that deploying the mechanism network-wide would significantly increase the number of model parameters, this study adopts a modular integration strategy by selectively embedding it after the 11th, 12th, and 13th inverted residual modules. This approach greatly improves the ability of the main network to extract features without compromising computational efficiency, resulting in a significant improvement in key performance metrics such as feature representation, object detection, and localization accuracy.

Experimental results and analysis

Datasets and hardware environment

The experimental data in this study were collected through a self-built platform, as shown in Fig. 8. An OpenMV H7 Plus device was used to acquire images at different water levels during the perfusion process, resulting in a dataset of 900 images across five water level intervals. All images were annotated using the Labelme tool. The data is partitioned into training, validation, and test sets in proportions of 80%, 10%, and 10%, respectively. To mitigate the risk of model overfitting, a dynamic data augmentation strategy is implemented during training, incorporating randomized operations including but not limited to rotation, brightness adjustment, and color transformation. Data augmentation is dynamically applied each time a batch is sampled from the dataset, with distinct augmentation operations stochastically applied to the original data per epoch. This strategy enhances the model’s generalization capability by exposing it to diversified feature patterns throughout training. Table 1 summarizes the hardware specifications and parameter settings of the testing platform.

Fig. 8.

Data collection and experimental platform.

Table 1.

Experimental platform hardware environment and parameters.

Name	Parameter
CPU	Intel Xeon Platinum
GPU	RTX 2080 Ti(11GB)
Operating system	Windows10
CUDA	10.0
Language	Python3.8
Input image resolution	300300
Initial learning rate	0.0001
Batch size	16
Epoch	200

Model evaluation criteria

To assess the enhanced network’s effectiveness, we employed three primary indicators: mean average precision (mAP), model parameter count (Params), and processing speed measured in FPS. The mAP metric measures the area beneath the curve that plots precision versus recall, capturing the balance between detection accuracy and coverage. Equations (3), (4), and (5) respectively define precision, recall, and mAP mathematically.

3

4

5

where TP refers to the number of samples correctly identified by the model, FP indicates the number of samples wrongly predicted, FN represents the count of missed samples belonging to negative categories, N stands for the total number of categories, and denotes the average precision for the i category.

Experimental results

The loss function of the SSD algorithm comprises classification loss and localization loss. Figure 9a illustrates that the proposed model achieves better convergence performance and reaches a lower final loss than the MobileNetV2-SSD baseline. Additionally, the enhanced architecture exhibits stronger data-fitting capacity and enhanced robustness.

Fig. 9.

Training loss and Precision-Recall curves before and after model improvement.

The Precision–Recall (P–R) curve is utilized to evaluate the classification performance across varying thresholds, where a larger area under the curve (Average Precision, AP) reflects superior model capability. From Fig. 9b, it can be observed that the MobileNetV2-SSD model experiences a significant decline in precision when recall is high, reflecting a higher false positive rate. The improved model avoids performance collapse in the high recall region and has a higher overall AUC-PR value, verifying the effectiveness of the improved strategy.

Ablation experiment

Table 2 summarizes the innovative model combinations based on different module configurations, while Table 3 presents the corresponding ablation study outcomes. The category-wise AP values for each model are illustrated in Fig. 10. Each module provides a beneficial impact on detection accuracy, and the approach introduced in this paper consistently outperforms others across all AP metrics. Although Model 1 exhibits lower detection accuracy when utilizing MobileNetV2 for feature extraction, it achieves an approximate 82.67% reduction in parameter quantity. Model 2 incorporates the Triplet Attention mechanism into the structure of Model 1, achieving a 2.75% improvement in mAP with only a 0.02% increase in parameter count, demonstrating that refined feature interaction design enables more cost-effective performance gains. Model 3 improves the mAP index by 5.80% over Model 1 by integrating the MS-DIRB module, which not only effectively compensates for the performance loss caused by the lightweight backbone network but also markedly increases the model’s capacity for capturing complex scene features. The experimental results indicate that MS-DIRB, as the core innovative module, contributes most significantly to performance optimization.

Table 2.

Schemes of innovations.

Model	VGG16	MobileNetV2	MS-DIRB	Triplet attention
SSD
Model 1
Model 2
Model 3
MD-MobileNetV2-SSD

Table 3.

Results of the ablation experiment.

Model	Precision (%)	Recall (%)	mAP(%)	Params
SSD	93.65	80.61	93.08	24280556
Model 1	89.12	69.54	89.08	4207624
Model 2	91.49	86.16	91.83	4208524
Model 3	92.75	90.03	94.88	4471640
MD-MobileNetV2-SSD	96.79	93.31	96.65	4472540

Fig. 10.

AP values of the ablation models at different liquid levels.

Comparison experiment

In order to validate the effectiveness of the proposed approach, a comprehensive comparison was made with mainstream object detection algorithms such as Faster R-CNN, SSD, YOLOv5s, YOLOv8s, YOLOv8n and MobileNetV2-SSD. Experimental results confirm that the proposed model delivers the most accurate performance in detecting liquid levels and barrel scale lines. The performance parameters of each algorithm are shown in Table 4, and the AP values for different liquid levels are shown in Fig. 11. As shown in Table 4, MobileNetV2-SSD significantly compresses parameter volume, reducing it to nearly one-sixth of that in SSD300, which achieves significant compression of the model volume while the inference speed is improved by 46.61%, but its detection accuracy suffers from an obvious compromise. To mitigate the loss in accuracy, the MD-MobileNetV2-SSD object detection algorithm is proposed in this study. The proposed method improves detection accuracy by 7.57% compared with MobileNetV2-SSD through only a 6.20% increase in parameter count, while achieving a 3.57% mAP gain compared with the SSD300 model and maintaining 35.89% faster inference speed.

Table 4.

Performance comparison of different models.

Model	Backbone	Precision (%)	Recall (%)	mAP (%)	Params	FPS
Faster R-CNN	VGG-16	95.48	89.33	94.32	142562335	8.90
SSD300	VGG-16	93.65	80.61	93.08	24280556	40.18
YOLOv5s	CSPDarknet53	92.70	84.49	93.46	7114785	46.64
YOLOv8s	C2f-Darknet	94.74	95.75	96.34	11166560	42.15
YOLOv8n	C2f-Darknet	93.70	95.75	94.26	3157200	59.04
MobileNetV2-SSD	MobileNetV2	89.12	69.54	89.08	4207624	58.91
MD-MobileNetV2-SSD	MobileNetV2	96.79	93.31	96.65	4472540	54.60

Fig. 11.

AP values of different models at different liquid levels.

In the comparative analysis with current mainstream YOLO series algorithms, the proposed method also demonstrates significant advantages. Compared to the YOLOv8s algorithm, our model achieves a 0.31% higher mAP while reducing the parameter count to only 40.05% of its counterpart, alongside a 29.54% increase in inference speed, thereby realizing an optimization in both accuracy and efficiency. When compared to the extremely lightweight YOLOv8n, our algorithm attains a 2.39% improvement in mAP at the cost of only a 7.52% decrease in speed, while exhibiting more balanced and stable performance on key categories, indicating a superior trade-off between lightness and detection capability. Furthermore, against the widely adopted YOLOv5s, our proposed algorithm shows comprehensive superiority across all three metrics: accuracy, speed, and model complexity. As illustrated in Fig. 11, across various liquid level scenarios, the proposed model achieves optimal detection performance on most key categories while maintaining its lightweight advantage, proving its robustness and practical utility in industrial settings. Experimental evidence confirms the superiority of the proposed model within the industrial context, while showing that the lightweight architecture enables efficient inference without substantially compromising accuracy.

To examine the performance of our proposed algorithm, experiments were conducted on the PASCAL VOC2007 dataset. Although this study focuses on industrial scenarios, the public dataset was selected to evaluate the model’s generalization capability, preventing the improved solution from overfitting to the specific industrial dataset and thereby comprehensively validating the effectiveness of the proposed enhancements. All tests were performed under identical hardware configurations to ensure implementation consistency. According to Table 5, the results indicate that the original MobileNetV2-SSD framework attains an mAP score 2.72% lower than the SSD300 baseline model for object detection tasks, primarily due to inherent limitations of its lightweight architecture. Our improved MobileNetV2-SSD achieves 63.69% mAP on the VOC2007 dataset, thereby narrowing the gap with SSD300 to merely 0.54% and significantly reducing computational costs. This improvement confirms that our model maintains its lightweight nature while approaching the detection accuracy of SSD300. These results validate the effectiveness of the improved solution in general object detection tasks and provide a reliable solution for balancing detection accuracy and computational efficiency.

Table 5.

Performance comparison between the original and improved models on the PASCAL VOC2007 dataset.

Model	Backbone	Precision (%)	Recall (%)	mAP (%)	Params	FPS
SSD300	VGG-16	72.94	56.79	64.23	26285486	38.87
MobileNetV2-SSD	MobileNetV2	77.04	51.75	61.51	6204004	52.59
MD-MobileNetV2-SSD	MobileNetV2	79.73	52.68	63.69	6468920	48.23

Experimental results presentation

To visually verify the performance of the algorithm, a visual comparative analysis of the test set samples is performed based on the confidence threshold of 0.5, and Fig. 12 presents the test results. The SSD300 detection algorithm suffers from defects such as leakage detection, localization bias, and insufficient confidence in detecting the low liquid level region. In the MobileNetV2-SSD algorithm, the same problems of leakage detection, false detection, and low confidence exist. The algorithm developed in this study effectively addresses the aforementioned issues, achieving a confidence level of over 90% in liquid detection during the filling process and an average accuracy of 87.2% in barrel scale recognition. It significantly enhances both localization precision and detection accuracy, demonstrating strong technical feasibility in industrial-grade dynamic liquid level monitoring scenarios.

Fig. 12.

Detection results of different algorithms for liquid level measurement.

A visual analysis was performed on the PASCAL VOC2007 dataset via randomized sampling of five test images, and the comparative detection results are depicted in Fig. 13. The proposed model achieves detection accuracy comparable to the SSD300 benchmark while requiring substantially lower computational complexity. Compared to the MobileNetV2-SSD architecture, our framework demonstrates significant enhancements in classification accuracy and bounding box localization precision, coupled with markedly reduced missed detection rates in high-density scenarios. These improvements effectively resolve the inherent limitations of MobileNetV2-SSD regarding constrained receptive fields and insufficient hierarchical feature representation. The achieved balance between detection performance and operational efficiency establishes a novel paradigm for deploying high-accuracy visual recognition systems in resource-limited environments.

Fig. 13.

Detection results of different algorithms on the PASCAL VOC2007 dataset.

Conclusion

This study proposes a lightweight object detection architecture for industrial perfusion scenarios, which achieves a synergistic improvement of detection efficiency and accuracy through multilevel structural optimization. A lightweight backbone network is used to significantly compress the model volume. A MS-DIRB is designed to establish multi-receptive-field feature interaction mechanism by integrating dilated convolutions with varying dilation rates. This effectively mitigates the fine-grained information loss during downsampling and enhances the feature representation capability for complex targets. To address the loss in accuracy introduced by the lightweight design, this work integrates Triplet Attention, which enhances the activation of key features by facilitating cross-channel communication. This addition incurs only a 0.02% parameter increase. Experimental results indicate that the improved approach achieves a notable enhancement in detection performance without compromising computational efficiency, making it well-suited for deployment in edge-embedded systems. The scheme effectively meets the real-time demand of dynamic liquid level monitoring tasks through multi-scale feature fusion and attention co-optimization mechanisms and provides a technical path with engineering landing value for liquid level monitoring. Future work will focus on two dimensions: model enhancement and engineering deployment. At the model level, the generalization capability will be validated through the construction of more complex industrial datasets, alongside continued exploration of advanced lightweight backbone networks and attention mechanisms to boost detection accuracy and efficiency in complex scenarios. At the engineering level, efforts will prioritize practical application, leveraging key techniques such as pruning and quantization to accomplish deployment and performance verification on edge computing platforms, ultimately achieving genuine industrial real-time monitoring.

Author contributions

X.Q. was responsible for experimental execution, manuscript writing, and figure organization, while L.J. was responsible for developing the initial proposal and experimental design, manuscript editing, and review.

Funding

This research received no external funding.

Data availability

Data is provided within the manuscript or supplementary information files.

Declarations

Competing interests

The authors declare no competing interests.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-29277-8.