Lightweight EG target detection system suitable for underwater navigation equipment

Abstract

The invention discloses a lightweight EG target detection system suitable for underwater navigation equipment. The lightweight EG target detection system comprises an image acquisition and preprocessing module, a lightweight backbone EGNet module, an A-PANet feature enhancement module and an output module. A seabed image is obtained and preprocessed through the image acquisition and preprocessing module; extracting a multi-scale feature map of the preprocessed seabed image by using a lightweight backbone EGNet; an APANet feature enhancement module is adopted to carry out adaptive multi-scale feature fusion on the multi-scale feature map so as to enhance the feature expression ability; and finally generating a target detection result through an output module. According to the method, the calculation complexity and the model parameter quantity are remarkably reduced while the high detection precision is kept, and the method is suitable for real-time target detection tasks under the resource limited environment of underwater navigation equipment and the like.

Description

Technical Field

[0001] This invention relates to the field of underwater target detection, and more particularly to a lightweight EG target detection system suitable for underwater navigation equipment. Background Technology

[0002] In the exploration and exploitation of marine resources, underwater target detection based on optical imaging has great potential. Many researchers have studied this and achieved good results, enabling underwater target detection to play a significant role in fisheries, resource exploration, underwater archaeology, and the exploration of marine ecosystems. However, underwater target detection technology faces multiple challenges, especially in complex underwater environments. Problems such as poor lighting conditions, blurry images, low contrast, and severe noise interference make it difficult for general target detection models to work effectively.

[0003] In existing technologies, the Boosting R-CNN method, through uncertainty modeling and hard sample mining, improves the detection performance of blurry and low-contrast images to some extent, but it does not fully address the unique optical characteristics of underwater environments. The color conversion network method converts color images to grayscale to alleviate underwater color absorption, but this method only addresses image color and does not consider the combined effects of image blur and noise interference, and it adds additional computational complexity. The FERNet method enhances feature representation capabilities through a composite connection backbone and receptive field enhancement module, addressing scale variations and sample imbalance, but it does not effectively handle underwater image blur and noise interference. The improved YOLOv7 algorithm uses the SPPFCSPC module to optimize computational efficiency, improving inference speed, but it does not solve the fundamental problem of poor underwater image quality. The SWIPENET backbone combined with the CMA training paradigm can generate high-resolution Hyper Feature Maps and gradually learn from noisy data, but this method still suffers from high computational complexity, insufficient adaptability to specific underwater optical characteristics, and does not fully consider the comprehensive processing of multi-dimensional image quality issues. Existing technical solutions generally suffer from the following problems when dealing with multiple issues such as underwater image blurring, low contrast, and noise interference: First, the network structure is complex, the number of parameters is large, and the computational load is high, making it unsuitable for real-time underwater embedded devices; Second, image features are affected by color shifts and low contrast, leading to a decrease in the model's feature extraction capability; Third, multi-scale target feature fusion is insufficient, resulting in low detection accuracy for small targets and the model training is prone to getting trapped in local optima, thus failing to fully utilize detection performance. Summary of the Invention

[0004] This invention provides a lightweight EG target detection system suitable for underwater navigation equipment, overcoming the problems that existing network structures are not suitable for real-time underwater embedded devices; image features are affected by multiple factors, leading to a decrease in model feature extraction capabilities; multi-scale target feature fusion is insufficient, resulting in low accuracy of small target detection and the model training is prone to getting trapped in local optima.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: A lightweight EG target detection system suitable for underwater navigation equipment includes: an image acquisition and preprocessing module, a lightweight backbone EGNet module, an A-PANet feature enhancement module, and an output module; The image acquisition and preprocessing module is used to acquire seabed images and perform preprocessing to obtain preprocessed seabed images, which are then output to the lightweight backbone EGNet module. The lightweight backbone EGNet module is used to extract features from the preprocessed seabed image to obtain multi-scale feature maps, and output them to the A-PANet feature enhancement module. The A-PANet feature enhancement module is used to adaptively fuse multi-scale feature maps to obtain a fused enhanced feature map, and then output it to the output module. The output module is used to generate the final target detection result based on the multi-scale fusion enhanced feature map.

[0006] Furthermore, the lightweight backbone EGNet module includes an initial feature extraction submodule, a first-scale feature generation submodule, a second-scale feature generation submodule, a third-scale feature generation submodule, and an output submodule; The initial feature extraction submodule consists of four CBS units and a first ELAN structure connected in sequence. It is used to extract initial features from the preprocessed seabed image and generate an initial feature map. The first scale feature generation submodule consists of a first MPConv block and a second ELAN structure. It is used to perform downsampling operation on the initial feature map through the first MPConv block to generate a first intermediate feature map; and input the first intermediate feature map into the second ELAN structure for feature enhancement operation to generate a small-scale feature map, which is then used as the first scale feature map. The second scale feature generation submodule consists of a second MPConv block and a first GhostBottleneck unit. It is used to perform downsampling operation on the first scale feature map through the second MPConv block to generate a second intermediate feature map; and input the second intermediate feature map into the first GhostBottleneck unit for lightweight feature extraction to generate a medium-scale feature map, which is then used as the second scale feature map. The third-scale feature generation submodule consists of a third MPConv block and a second GhostBottleneck unit. It is used to perform downsampling operation on the second-scale feature map through the third MPConv block to generate a third intermediate feature map; and input the third intermediate feature map into the second GhostBottleneck unit for lightweight feature extraction to generate a large-scale feature map, which is then used as the third-scale feature map. The output submodule is used to construct a multi-scale feature representation from the first-scale feature map, the second-scale feature map, and the third-scale feature map, and output it to the A-PANet feature enhancement module.

[0007] Furthermore, the A-PANet feature enhancement module includes a multi-scale feature processing submodule, a first output fusion feature generation submodule, a second output fusion feature generation submodule, a third output fusion feature generation submodule, and a RepConv submodule; The multi-scale feature processing submodule consists of two convolutional layers and an SPPF block. It is used to perform convolution operations on the first-scale feature map through the first convolutional layer to obtain the first processed feature; to perform convolution operations on the second-scale feature map through the second convolutional layer to obtain the second processed feature; and to perform SPPF operations on the third-scale feature map through the SPPF block to obtain the third processed feature. The first output fusion feature generation submodule consists of two upsampling layers, two ELAN structures, and two corresponding cascaded operations. It is used to upsample the third processed feature through the first upsampling layer, and to perform channel cascaded operation with the second processed feature through the first cascaded layer to obtain the first fusion feature; and to perform feature enhancement operation on the first fusion feature through the first ELAN structure to generate the first fusion enhanced feature. The first fusion enhancement feature is upsampled through the second upsampling layer and then channel-concatenated with the first processed feature through the second cascade layer to obtain the second fusion feature. The second fusion feature is then enhanced through the second ELAN structure to generate the second fusion enhancement feature, which is then output as the first output fusion feature to the RepConv submodule. The second output fusion feature generation submodule consists of a first MP2 layer, a first AFF block, and a third ELAN structure. It is used to perform downsampling operation on the first output fusion feature through the first MP2 layer to obtain a first intermediate fusion feature; to perform adaptive fusion of the first intermediate fusion feature and the first fusion enhancement feature through the first AFF block to obtain a first adaptive fusion feature; and to perform feature enhancement operation on the first adaptive fusion feature through the third ELAN structure to generate a third fusion enhancement feature, which is then output as the second output fusion feature to the RepConv submodule. The third output fusion feature generation submodule consists of a second MP2 layer, a second AFF block, and a fourth ELAN structure. It is used to perform a downsampling operation on the second output fusion feature through the second MP2 layer to obtain a second intermediate fusion feature; to perform adaptive fusion of the second intermediate fusion feature and the third processed feature through the second AFF block to obtain a second adaptive fusion feature; and to perform feature enhancement operation on the second adaptive fusion feature through the fourth ELAN structure to generate a fourth fusion enhancement feature, which is then output as the third output fusion feature to the RepConv submodule. The RepConv submodule is used to fuse the first output fusion feature, the second output fusion feature and the third output fusion feature to obtain a fusion enhanced feature map and output it to the output module.

[0008] Furthermore, the AFF block is a multi-scale channel attention mechanism; the multi-scale channel attention mechanism includes global branches and local branches; The global branch obtains global channel attention based on the input feature map through global average pooling, thereby obtaining global features; The local branch obtains local channel attention through multiple parallel average pooling layers with different pooling kernel sizes; the local branch can be represented as:

[0009] In the formula, For the output of a local branch; Input feature map; For activation functions; and For pointwise convolution; Based on global and local branches, the multi-scale channel attention mechanism can be represented as:

[0010] In the formula, For the output of the multi-scale channel attention mechanism; It is the sigmoid activation function; This is the output of the global branch; Element-wise multiplication; For cascading operations; Based on the multi-scale channel attention mechanism, the operation of the AFF module is described as follows:

[0011] In the formula, Z is the output of the AFF module; This is a multi-scale channel attention mechanism; X and Y are input features.

[0012] Furthermore, the lightweight EG target detection system also includes a secondary training module; The secondary training module includes a first training sub-module and a second training sub-module; The first training submodule is used to perform freeze training and unfreeze training on the lightweight EG target detection system to obtain the first training lightweight EG target detection system, and save the weight with the highest average accuracy index during the training process as the pre-training weight. The second training submodule is used to initialize the parameters of the lightweight EG object detection system trained in the first training according to the pre-trained weights, and to perform a second freeze training and unfreeze training, so as to finally obtain the trained lightweight EG object detection system.

[0013] Beneficial effects: This invention provides a lightweight EG target detection system suitable for underwater navigation equipment. By adopting a lightweight backbone EGNet, the number of model parameters and computational complexity are significantly reduced. This enables the system to operate efficiently in environments with limited computing resources, such as embedded devices, meeting the urgent need of underwater navigation equipment for real-time target detection. The A-PANet feature enhancement module effectively aggregates deep semantic information and shallow detail information through an adaptive, bidirectional, multi-scale feature fusion mechanism. This significantly improves the system's feature representation capability for underwater targets (especially small, blurred, and occluded targets), thus maintaining high detection accuracy and robustness in underwater scenes with uneven lighting, turbid water, and complex backgrounds. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the target detection system of the present invention; Figure 2 This is a data flow diagram of the target detection system in an embodiment of the present invention; Figure 3 This is a structural diagram of the lightweight backbone EGNet module in an embodiment of the present invention; Figure 4 This is a structural diagram of the GhostBottleneck unit in an embodiment of the present invention; Figure 5 This is a structural diagram of the AFF block in an embodiment of the present invention; Figure 6This is a structural diagram of the multi-scale channel attention mechanism in an embodiment of the present invention; Figure 7 This is a comparison diagram between the present invention and existing detectors in this embodiment; Figure 8 This is a qualitative comparison result of the URPC2019 dataset in an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] This embodiment provides a lightweight EG target detection system suitable for underwater navigation equipment, such as Figure 1 As shown, it includes: an image acquisition and preprocessing module, a lightweight backbone fusion high-efficiency layer - Phantom Network (ELAN-GhostNet, EGNet) module, a feature fusion - Path Aggregation Network (AFF-Path Aggregation Network, A-PANet) feature enhancement module, and an output module; The image acquisition and preprocessing module is used to acquire seabed images and perform preprocessing to obtain preprocessed seabed images, which are then output to the lightweight backbone EGNet module. The lightweight backbone EGNet module is used to extract features from the preprocessed seabed image to obtain multi-scale feature maps, and output them to the A-PANet feature enhancement module. The A-PANet feature enhancement module is used to adaptively fuse multi-scale feature maps to obtain a fused enhanced feature map, and then output it to the output module. The output module is used to generate the final target detection result based on the multi-scale fusion enhanced feature map.

[0018] Specifically, such as Figure 2As shown, the system first acquires raw seabed images through an image acquisition and preprocessing module, and then performs standardized processing such as color correction, denoising, and enhancement to overcome underwater ambient light attenuation, scattering, and noise interference, outputting high-quality preprocessed images that lay a stable foundation for subsequent deep feature extraction. Secondly, the preprocessed images are input into a lightweight backbone EGNet module, which extracts multi-scale feature maps with rich semantic information from different network layers, effectively balancing detection accuracy with the resource constraints of embedded device deployment. Thirdly, the multi-scale feature maps are input into the A-PANet feature enhancement module, which aggregates deep semantic features and shallow detail features from top to bottom and bottom to top, achieving adaptive fusion and enhancement of contextual information for targets at different scales (especially small and blurry targets), generating more discriminative fused and enhanced feature maps. Finally, the output module, based on this fused and enhanced feature map, performs calculations using a preset detection head (such as anchor box regression and classifier) ​​to accurately locate and identify the target category and location in the seabed image, generating the final target detection result.

[0019] Specifically, the preprocessing of the image acquisition and preprocessing module includes image size normalization, noise filtering, and color space conversion; the image size normalization, noise filtering, and color space conversion are existing technologies and will not be elaborated here.

[0020] Preferred, such as Figure 3 As shown, the lightweight backbone EGNet module integrates the multi-branch feature extraction advantages of the EfficientLayer Aggregation Network (ELAN) structure with the low parameter characteristics of the Ghostbottleneck module; while significantly reducing the number of model parameters and computational complexity, it can still effectively maintain high target detection accuracy. The lightweight backbone EGNet module includes an initial feature extraction submodule, a first-scale feature generation submodule, a second-scale feature generation submodule, a third-scale feature generation submodule, and an output submodule. The initial feature extraction submodule consists of four CBS units and a first ELAN structure connected in sequence. It is used to extract initial features from the preprocessed seabed image and generate an initial feature map. The first scale feature generation submodule consists of a first MPConv block and a second ELAN structure. It is used to perform downsampling operation on the initial feature map through the first MPConv block to generate a first intermediate feature map; and input the first intermediate feature map into the second ELAN structure for feature enhancement operation to generate a small-scale feature map, which is then used as the first scale feature map. The second scale feature generation submodule consists of a second MPConv block and a first GhostBottleneck unit. It is used to perform downsampling operation on the first scale feature map through the second MPConv block to generate a second intermediate feature map; and input the second intermediate feature map into the first GhostBottleneck unit for lightweight feature extraction to generate a medium-scale feature map, which is then used as the second scale feature map. The third-scale feature generation submodule consists of a third MPConv block and a second GhostBottleneck unit. It is used to perform downsampling operation on the second-scale feature map through the third MPConv block to generate a third intermediate feature map; and input the third intermediate feature map into the second GhostBottleneck unit for lightweight feature extraction to generate a large-scale feature map, which is then used as the third-scale feature map. The output submodule is used to construct a multi-scale feature representation from the first-scale feature map, the second-scale feature map, and the third-scale feature map, and output it to the A-PANet feature enhancement module.

[0021] The lightweight backbone EGNet module includes an initial feature extraction submodule, a first-scale feature generation submodule, a second-scale feature generation submodule, a third-scale feature generation submodule, and an output submodule. The initial feature extraction submodule consists of four convolutional-batch normalization-activation function (Conv-BN-SiLU, CBS) units connected in sequence and a first ELAN structure. It is used to extract initial features from the preprocessed seabed image (640×640×3) to generate an initial feature map. The CBS unit is composed of a convolutional layer, a batch normalization layer, and a SiLU activation function connected in series. The first scale feature generation submodule consists of a first max pooling-convolution (MPConv) block and a second ELAN structure. It is used to perform downsampling operation on the initial feature map through the first MPConv block to generate a first intermediate feature map; and input the first intermediate feature map into the second ELAN structure for feature enhancement operation to generate a small-scale feature map (80×80×256), which is used as the first scale feature map. The MPConv is a downsampling module that includes a max pooling layer and a convolutional layer; The second scale feature generation submodule consists of a second MPConv block and a first GhostBottleneck unit. It is used to perform downsampling operation on the first scale feature map through the second MPConv block to generate a second intermediate feature map; and input the second intermediate feature map into the first GhostBottleneck unit for lightweight feature extraction to generate a medium-scale feature map (40×40×512), which is used as the second scale feature map. The GhostBottleneck unit is a residual bottleneck structure built based on the lightweight concept of GhostNet; The third-scale feature generation submodule consists of a third MPConv block and a second GhostBottleneck unit. It is used to perform downsampling operation on the second-scale feature map through the third MPConv block to generate a third intermediate feature map; the third intermediate feature map is input into the second GhostBottleneck unit for lightweight feature extraction to generate a large-scale feature map (20×20×1024), which is used as the third-scale feature map. The output submodule is used to construct multi-scale features from the first-scale feature map, the second-scale feature map, and the third-scale feature map, and output them to the A-PANet feature enhancement module.

[0022] In this embodiment, the ELAN structure enables the network to continuously enhance its learning ability without disrupting the original gradient paths. By considering the shortest and longest gradient paths at each layer and the longest gradient path across the entire network, it allows the network to stack more computational blocks without reducing parameter usage. like Figure 4 As shown, the GhostBottleneck unit consists of two stacked ghost modules and uses skip connections to alleviate the gradient vanishing problem of the ELAN structure. The execution steps of the Ghost module are as follows: First, perform regular convolution on the second intermediate feature map to obtain a feature map with half the number of channels; Secondly, a depthwise convolution is performed on the feature map with half the number of channels to obtain a depth feature map; Finally, the feature map with half the number of channels is concatenated with the depth feature map to obtain the mesoscale feature map; The advantage of the Ghost module is that it can achieve the same output as regular convolution with less computation and fewer parameters, thereby improving efficiency and making the model lightweight.

[0023] Preferably, the A-PANet feature enhancement module removes redundant convolutional layers that have little impact on detection performance through model pruning techniques, and introduces a lightweight adaptive feature fusion (AFF) path aggregation network (PANet) into the top-down feature fusion path; by enhancing feature information through top-down and bottom-up feature fusion, features at different levels can be effectively fused, information flow can be enhanced, and the accuracy of target detection can be improved; The A-PANet feature enhancement module includes a multi-scale feature processing submodule, a first output fusion feature generation submodule, a second output fusion feature generation submodule, a third output fusion feature generation submodule, and a RepConv submodule; The multi-scale feature processing submodule consists of two convolutional layers and a Spatial Pyramid Pooling Fast (SPPF) block. It is used to perform convolution operations on the feature map at the first scale through the first convolutional layer to obtain the first processed feature; to perform convolution operations on the feature map at the second scale through the second convolutional layer to obtain the second processed feature; and to perform SPPF operations on the feature map at the third scale through the SPPF block to obtain the third processed feature. The SPPF block consists of at least two cascaded max pooling layers of the same kernel size, used to concatenate the original input with all pooling outputs in the channel dimension; The first output fusion feature generation submodule consists of two upsampling layers, two ELAN structures, and two corresponding cascaded operations. It is used to upsample the third processed feature through the first upsampling layer, and to perform channel cascaded operation with the second processed feature through the first cascaded layer to obtain the first fusion feature; and to perform feature enhancement operation on the first fusion feature through the first ELAN structure to generate the first fusion enhanced feature. The first fusion enhancement feature is upsampled through the second upsampling layer and channel-concatenated with the first processed feature through the second cascade layer to obtain the second fusion feature. The second fusion feature is then enhanced through the second ELAN structure to generate the second fusion enhancement feature, which is then output as the first output fusion feature to the reparameterizable convolution (RepConv) submodule. The second output fusion feature generation submodule consists of a first max pooling (MP) 2 layer, a first AFF block, and a third ELAN structure. It is used to downsample the first output fusion feature through the first MP2 layer to obtain a first intermediate fusion feature; to adaptively fuse the first intermediate fusion feature and the first fusion enhancement feature through the first AFF block to obtain a first adaptive fusion feature; and to perform feature enhancement on the first adaptive fusion feature through the third ELAN structure to generate a third fusion enhancement feature, which is then output as the second output fusion feature to the RepConv submodule. The third output fusion feature generation submodule consists of a second MP2 layer, a second AFF block, and a fourth ELAN structure. It is used to perform a downsampling operation on the second output fusion feature through the second MP2 layer to obtain a second intermediate fusion feature; to perform adaptive fusion of the second intermediate fusion feature and the third processed feature through the second AFF block to obtain a second adaptive fusion feature; and to perform feature enhancement operation on the second adaptive fusion feature through the fourth ELAN structure to generate a fourth fusion enhancement feature, which is then output as the third output fusion feature to the RepConv submodule. The RepConv submodule is used to fuse the first output fusion feature, the second output fusion feature and the third output fusion feature to obtain a fusion enhanced feature map and output it to the output module.

[0024] Preferably, the AFF block is a multi-scale channel attention mechanism; the multi-scale channel attention mechanism includes global branches and local branches; The global branch obtains global channel attention based on the input feature map through global average pooling, thereby obtaining global features; like Figure 6 As shown, the local branch obtains local channel attention through multiple parallel average pooling layers with different pooling kernel sizes; the local branch can be represented as:

[0025] In the formula, For the output of a local branch; Input feature map; For activation functions; and For pointwise convolution; Based on global and local branches, the multi-scale channel attention mechanism can be represented as:

[0026] In the formula, For the output of the multi-scale channel attention mechanism; It is the sigmoid activation function; This is the output of the global branch; Element-wise multiplication; For cascading operations; like Figure 5 As shown, based on the multi-scale channel attention mechanism, the operation of the AFF module is described as follows:

[0027] In the formula, Z is the output of the AFF module; This is a multi-scale channel attention mechanism; X and Y are input features.

[0028] In this embodiment, by introducing the AFF mechanism, A-PANet can more effectively handle targets at different scales and semantic differences between feature maps; this allows the entire model to capture and utilize local and global information in images more efficiently while maintaining its lightweight nature, thereby enhancing the model's feature representation capabilities and improving the detection accuracy of underwater targets.

[0029] Preferably, the lightweight EG target detection system further includes a secondary training module; The secondary training module includes a first training sub-module and a second training sub-module; The first training submodule is used to perform freeze training and unfreeze training on the lightweight EG target detection system to obtain the first training lightweight EG target detection system, and save the weight with the highest average accuracy index during the training process as the pre-training weight. The second training submodule is used to initialize the parameters of the lightweight EG object detection system trained in the first training according to the pre-trained weights, and to perform a second freeze training and unfreeze training, so as to finally obtain the trained lightweight EG object detection system.

[0030] In this embodiment, weights trained on the large ImageNet dataset are used as the initial pre-trained weights. The total number of training iterations is set to 350, including 50 freeze training iterations and 300 unfreeze training iterations. During freeze training, the lightweight backbone EGNet module is frozen to prevent the initial backbone weights from being destroyed in the early stages of training, and its gradient is not updated. The batch size for freeze training is set to 8, the batch size for unfreeze training is set to 4, and num-works is set to 4 to allow the network to read data in multiple threads. In the second training in this embodiment, the weights with the highest accuracy in the first training (based on the mean accuracy index) are saved as the pre-trained model, and then a new round of training is started. In the second training process, an early stopping method is used to avoid model overfitting.

[0031] In the first embodiment, the performance of the target detection system in this embodiment is verified using the open-source dataset URPC2019; The open-source dataset URPC2019 contains 4757 images, covering four types of organisms: sea urchins, starfish, sea cucumbers, and scallops. Since the test set of the URPC2019 dataset is no longer available, the ratio of the sum of the training set and the validation set to the test set is set to 9:1, and the ratio of the training set to the validation set is also set to 9:1. Performance verification metrics include model lightweighting evaluation metrics and detection accuracy evaluation metrics; wherein, the model lightweighting evaluation metrics are parameterization metrics, billion floating-point operations per second (GFLOPS), and frame rate; the detection accuracy evaluation metrics are average precision (…). AP ) and average accuracy at an IoU threshold of 0.5 ( mAP 50 ); The AP reflects the average recognition accuracy for each category. The calculation formula for AP needs to incorporate precision and recall, where recall is the horizontal axis of the coordinate system and precision is the vertical axis. The AP can be obtained by integrating the PR (precision-recall) curve. value; The mAP Reflecting the average recognition accuracy across all categories, it is one of the most important metrics in the field of object detection; In performance verification, comparisons were made with YOLOv7-tiny, BG-YOLO, YOLOX-s, YOLOv5-s, YOLOv8-n, YOLOv11-n, and Boosting R-CNN. The comparison results are shown in Table 1. Table 1. Comparison results with other detectors on the URPC2019 dataset.

[0032] Table 1 shows that the object detection system in this embodiment achieved the highest performance on the open-source dataset URPC2019. mAP Value; The target detection system in this embodiment improves upon the baseline model YOLOv7-tiny by 0.46%. mAPThe system achieves a 43.64% reduction in the number of parameters and a 23.08% reduction in GFLOPS consumption. Furthermore, while the target detection system in this embodiment is not optimal in terms of the number of parameters and GFLOPS, it only exceeds the suboptimal YOLOv8-n by 0.237M and 1.239G, respectively, and exhibits a significant accuracy advantage over YOLOv8-n. Compared to the latest detector YOLOv11-n, the target detection system in this embodiment still has an advantage in the number of parameters (only 0.809M more than YOLOv11-n) and significantly outperforms it in accuracy. mAP The metrics are 0.89% higher than YOLOv11-n; Table 2 shows the accuracy comparison results of the target detection system in this embodiment with other detectors across all categories. The target detection system in this embodiment not only... mAP It outperformed other detectors in terms of metrics and achieved the highest scores in both categories. AP The value ranked second in the other two categories; Table 2 shows the accuracy comparison results with other detectors across all categories.

[0033] Figure 7 The image shows the qualitative comparison results of the URPC2019 open-source dataset. As can be seen from the image, in the first column on the left, only the target detection system in this embodiment matches the true label, while other detectors either falsely detect or miss detections of sea urchins and starfish. In the second column, the comparison methods falsely detect sea cucumbers. In the third column, only the target detection system in this embodiment accurately detects all targets. The fourth column shows that all comparison methods miss detections of sea urchins, while the target detection system in this embodiment is completely consistent with the true label. The comparison results demonstrate that the target detection system in this embodiment can indeed detect more complete targets.

[0034] In the second embodiment, the generalization performance of the object detection system in this embodiment is verified using the open-source dataset URPC2020. The open-source dataset URPC2019 contains 5543 images, as well as more complex scenarios; In the generalization validation, comparisons were made with YOLOv7-tiny, BG-YOLO, YOLOX-s, YOLOv5-s, YOLOv8-n, YOLOv11-n, and Boosting R-CNN. The comparison results are shown in Table 3. Table 3 shows the experimental results comparing the detectors with other detectors on the URPC2020 dataset.

[0035] As shown in Table 3, the target detection system in this embodiment achieves optimal performance on the URPC2020 dataset. mAP Compared to the classic detector YOLOv5-s, the accuracy is improved by 1.37%, and compared to the advanced detector YOLOvX-tiny, it is improved by 0.32%. Meanwhile, the target detection system in this embodiment consistently leads in frame rate, fully demonstrating its superior real-time performance. Compared to Boosting R-CNN, which is optimized for underwater environments, the EG target detector significantly reduces the number of parameters by 44.716M while maintaining an accuracy advantage of 0.66%. Compared to the latest detector YOLOv11-n, the target detection system in this embodiment has only 0.809M more parameters, but... mAP The indicator is 2.06% ahead, which significantly highlights the excellent balance between performance and efficiency; Using the open-source dataset URPC2019, the accuracy comparison results of the object detection system in this embodiment with other detectors across all categories are shown in Table 4. Table 4. Accuracy comparison results with other detectors across all categories

[0036] As shown in Table 4, the target detection system in this embodiment ranks first in overall detection accuracy and leads in two sub-categories. Especially in the highly challenging category of sea cucumber, its accuracy is improved by 0.83% compared to the second-best solution, fully demonstrating the model's excellent performance in complex underwater scenarios. In general, increasing model complexity (such as deepening or widening neural networks) can capture more image target feature information, thereby improving detection accuracy; however, this also means requiring more computing resources (higher GFLOPS), because complex models need to perform more floating-point operations during training and inference. In underwater environments, model lightweighting is just as important as detection accuracy. As can be seen from the above comparative experiments, the target detection system in this embodiment balances detection accuracy and model lightweighting, not only significantly reducing the number of model parameters and GFLOPS consumption, but also improving detection accuracy to a certain extent. The above comparative experiments verify the effectiveness of the target detection system in this embodiment.

[0037] In the third embodiment, the effectiveness of the target detection system in this embodiment is quantitatively evaluated using the control variable method, and the rationality of each module is verified through experiments on the URPC2019 dataset; the experimental results are shown in Table 5: Table 5 Experimental Results

[0038] As shown in Table 5, EGNet, as a lightweight backbone network, successfully reduced the number of model parameters by 1.937M and the computational cost by 2.474G while slightly affecting accuracy and frame rate. The AFF module further improved the detection accuracy by 0.18% while maintaining the same number of parameters and computational efficiency. SPPF and pruning techniques effectively reduced the parameter size without sacrificing detection speed and accuracy by optimizing the model structure. Thanks to the synergistic optimization of these modules, the target detection system in this embodiment achieves comprehensive leadership in all performance indicators. To further verify the effectiveness of the lightweight backbone EGNet module and AFF module in the target detection system of this embodiment, in addition to quantitative experimental results, qualitative experimental analysis was also conducted in this embodiment; such as Figure 8 As shown, the original model falsely detected two sea urchins. After adding EGNet, the number of false detections decreased to 1. After adding the AFF module, the results were completely consistent with the real data. The experimental results show that EGNet and AFF not only achieve model lightweighting without reducing the underwater target detection accuracy, but also alleviate the underwater target false detection phenomenon to a certain extent. The target detection system in this embodiment has the best overall performance and achieves an excellent balance between model lightweighting and detection accuracy.

[0039] The present invention has the following beneficial effects: This invention provides a lightweight EG target detection system suitable for underwater navigation equipment. By adopting a lightweight backbone EGNet, the number of model parameters and computational complexity are significantly reduced. This enables the system to operate efficiently in environments with limited computing resources, such as embedded devices, meeting the urgent need of underwater navigation equipment for real-time target detection. The A-PANet feature enhancement module effectively aggregates deep semantic information and shallow detail information through an adaptive, bidirectional, multi-scale feature fusion mechanism. This significantly improves the system's feature representation capability for underwater targets (especially small, blurred, and occluded targets), thus maintaining high detection accuracy and robustness in underwater scenes with uneven lighting, turbid water, and complex backgrounds.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A lightweight EG target detection system suitable for underwater navigation equipment, characterized in that, include: Image acquisition and preprocessing module, lightweight backbone EGNet module, A-PANet feature enhancement module and output module; The image acquisition and preprocessing module is used to acquire seabed images and perform preprocessing to obtain preprocessed seabed images, which are then output to the lightweight backbone EGNet module. The lightweight backbone EGNet module is used to extract features from the preprocessed seabed image to obtain multi-scale feature maps, and output them to the A-PANet feature enhancement module. The A-PANet feature enhancement module is used to adaptively fuse multi-scale feature maps to obtain a fused enhanced feature map, and then output it to the output module. The output module is used to generate the final target detection result based on the multi-scale fusion enhanced feature map.

2. The lightweight EG target detection system for underwater navigation equipment according to claim 1, characterized in that, The lightweight backbone EGNet module includes an initial feature extraction submodule, a first-scale feature generation submodule, a second-scale feature generation submodule, a third-scale feature generation submodule, and an output submodule. The initial feature extraction submodule consists of four CBS units and a first ELAN structure connected in sequence. It is used to extract initial features from the preprocessed seabed image and generate an initial feature map. The first scale feature generation submodule consists of a first MPConv block and a second ELAN structure. It is used to perform downsampling operation on the initial feature map through the first MPConv block to generate a first intermediate feature map; and input the first intermediate feature map into the second ELAN structure for feature enhancement operation to generate a small-scale feature map, which is then used as the first scale feature map. The second scale feature generation submodule consists of a second MPConv block and a first GhostBottleneck unit. It is used to perform downsampling operation on the first scale feature map through the second MPConv block to generate a second intermediate feature map; and input the second intermediate feature map into the first GhostBottleneck unit for lightweight feature extraction to generate a medium-scale feature map, which is then used as the second scale feature map. The third-scale feature generation submodule consists of a third MPConv block and a second GhostBottleneck unit. It is used to perform downsampling operation on the second-scale feature map through the third MPConv block to generate a third intermediate feature map; and input the third intermediate feature map into the second GhostBottleneck unit for lightweight feature extraction to generate a large-scale feature map, which is then used as the third-scale feature map. The output submodule is used to construct a multi-scale feature representation from the first-scale feature map, the second-scale feature map, and the third-scale feature map, and output it to the A-PANet feature enhancement module.

3. The lightweight EG target detection system for underwater navigation equipment according to claim 1, characterized in that, The A-PANet feature enhancement module includes a multi-scale feature processing submodule, a first output fusion feature generation submodule, a second output fusion feature generation submodule, a third output fusion feature generation submodule, and a RepConv submodule; The multi-scale feature processing submodule consists of two convolutional layers and an SPPF block. It is used to perform convolution operations on the first-scale feature map through the first convolutional layer to obtain the first processed feature; to perform convolution operations on the second-scale feature map through the second convolutional layer to obtain the second processed feature; and to perform SPPF operations on the third-scale feature map through the SPPF block to obtain the third processed feature. The first output fusion feature generation submodule consists of two upsampling layers, two ELAN structures, and two corresponding cascaded operations. It is used to upsample the third processed feature through the first upsampling layer, and to perform channel cascaded operation with the second processed feature through the first cascaded layer to obtain the first fusion feature; and to perform feature enhancement operation on the first fusion feature through the first ELAN structure to generate the first fusion enhanced feature. The first fusion enhancement feature is upsampled through the second upsampling layer, and then channel-concatenated with the first processed feature through the second cascade layer to obtain the second fusion feature. The second ELAN structure is used to perform feature enhancement operations on the second fused feature to generate a second fused enhanced feature, which is then output as the first output fused feature to the RepConv submodule. The second output fusion feature generation submodule consists of a first MP2 layer, a first AFF block, and a third ELAN structure. It is used to perform a downsampling operation on the first output fusion feature through the first MP2 layer to obtain the first intermediate fusion feature. The first intermediate fusion feature and the first fusion enhancement feature are adaptively fused through the first AFF block to obtain the first adaptive fusion feature; The first adaptive fusion feature is enhanced by the third ELAN structure to generate the third fusion enhanced feature, which is then output as the second output fusion feature to the RepConv submodule. The third output fusion feature generation submodule consists of a second MP2 layer, a second AFF block, and a fourth ELAN structure. It is used to perform a downsampling operation on the second output fusion feature through the second MP2 layer to obtain the second intermediate fusion feature. The second intermediate fusion feature and the third processed feature are adaptively fused through the second AFF block to obtain the second adaptive fusion feature. The second adaptive fusion feature is enhanced by the fourth ELAN structure to generate the fourth fusion enhanced feature, which is then output as the third output fusion feature to the RepConv submodule. The RepConv submodule is used to fuse the first output fusion feature, the second output fusion feature and the third output fusion feature to obtain a fusion enhanced feature map and output it to the output module.

4. A lightweight EG target detection system suitable for underwater navigation equipment according to claim 1, characterized in that, The AFF block is a multi-scale channel attention mechanism; the multi-scale channel attention mechanism includes global branches and local branches; The global branch obtains global channel attention based on the input feature map through global average pooling, thereby obtaining global features; The local branch obtains local channel attention through multiple parallel average pooling layers with different pooling kernel sizes; the local branch can be represented as: In the formula, For the output of a local branch; Input feature map; For activation functions; and For pointwise convolution; Based on global and local branches, the multi-scale channel attention mechanism can be represented as: In the formula, For the output of the multi-scale channel attention mechanism; It is the sigmoid activation function; This is the output of the global branch; Element-wise multiplication; For cascading operations; Based on the multi-scale channel attention mechanism, the operation of the AFF module is described as follows: In the formula, Z is the output of the AFF module; This is a multi-scale channel attention mechanism; X and Y are input features.

5. A lightweight EG target detection system suitable for underwater navigation equipment according to claim 1, characterized in that, The lightweight EG target detection system also includes a secondary training module; The secondary training module includes a first training sub-module and a second training sub-module; The first training submodule is used to perform freeze training and unfreeze training on the lightweight EG target detection system to obtain the first training lightweight EG target detection system, and save the weight with the highest average accuracy index during the training process as the pre-training weight. The second training submodule is used to initialize the parameters of the lightweight EG object detection system trained in the first training according to the pre-trained weights, and to perform a second freeze training and unfreeze training, so as to finally obtain the trained lightweight EG object detection system.

Images