Underwater small target detection system based on double purification feature fusion and construction method thereof
The underwater small target detection system, which integrates dual purification features, solves the problems of underwater image degradation and feature fusion, achieving high accuracy and real-time performance in underwater target detection, and is suitable for embedded devices.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SANJIANG UNIVERSITY
- Filing Date
- 2026-01-19
- Publication Date
- 2026-06-05
AI Technical Summary
The complexity and unique characteristics of the underwater environment lead to severe image fogging. Existing technologies struggle to effectively address image degradation and feature fusion issues in underwater small target detection. Furthermore, high-precision deep learning modules are difficult to deploy on embedded devices, making it impossible to balance accuracy and real-time performance.
An underwater small target detection system based on dual-cleaning feature fusion is adopted, including an image preprocessing module, an improved Faster R-CNN network, and a multi-scale feature fusion mechanism. Combined with hardware acceleration and memory optimization, it achieves both high accuracy and embedded real-time performance in underwater target detection.
It significantly improves underwater image quality, enhances the accuracy of small target detection, and enables efficient real-time target detection on embedded platforms.
Smart Images
Figure CN122156942A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision and embedded artificial intelligence technology, specifically relating to an underwater small target detection system based on dual purification feature fusion and its construction method. Background Technology
[0002] Underwater small target detection plays a crucial role in marine resource exploration, autonomous underwater robot operations, and ecological environment monitoring. However, the complexity and unique characteristics of the underwater environment pose significant challenges to visual inspection technologies.
[0003] First, water has a strong absorption and scattering effect on light, resulting in severe fogging in the captured images. The fogging effect leads to an average decrease in image contrast. Although there are traditional image enhancement methods such as CLAHE, they often fail to physically restore the true colors and details of the scene.
[0004] Secondly, underwater targets of interest, such as many marine organisms, are usually small in size and irregular in shape. In deep learning-based target detection models, multi-scale feature fusion is a key technology. However, Feature Pyramid Network (FPN) is prone to semantic differences when directly fusing deep semantic features and shallow detail features. Spatial Pyramid Pooling (SPP) uses fixed-size grid pooling, which leads to the loss of details of irregular targets.
[0005] Finally, high-precision deep learning modules, such as the Transformer attention module, have high computational complexity and a large number of parameters, making them difficult to deploy directly on embedded devices where computing resources, storage space, and power consumption are severely limited, thus failing to meet the requirements for both accuracy and real-time detection.
[0006] Therefore, there is an urgent need for a target detection technology that can simultaneously solve underwater image degradation, improve the quality of small target feature fusion, and run efficiently on embedded platforms. Summary of the Invention
[0007] To address the problems existing in the prior art, this invention proposes an underwater small target detection system and its construction method based on dual-purification feature fusion. By using a multi-scale feature conflict filtering method in image preprocessing, it solves problems such as fogging interference, feature fusion conflict and loss of details, achieving a dual breakthrough in underwater target detection accuracy and embedded real-time performance.
[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0009] The underwater small target detection system based on dual-cleaning feature fusion includes: an image preprocessing module for preprocessing the input underwater image using an improved dark channel prior algorithm; and an underwater small target detection model, which is an improved Faster algorithm. The R-CNN network comprises a backbone network, a feature pyramid network, a context aggregation module, a multi-scale filtering module, a region proposal network, a ROI pooling layer, and a detection head. The backbone network uses the DarkNet53 architecture to extract primary feature maps from the preprocessed image. The feature pyramid network upsamples the primary feature maps output by the backbone network and fuses them with downsampled feature maps of the corresponding scale to generate multi-scale feature maps. The context aggregation module performs multi-scale pooling operations on the feature maps at each scale output by the feature pyramid network to enhance contextual information representation. The multi-scale filtering module employs a dual-cleaning fusion mechanism of channel cleansing and spatial cleansing to adaptively weight and fuse the multi-scale features output by the context aggregation module. The region proposal network generates candidate regions based on the dual-cleaned fused features, which are then classified and bounding box regressed by the ROI pooling layer and the detection head to output the final detection result. The underwater small target detection model is optimized and deployed on the Jetson TX2 platform using TensorRT to achieve real-time inference for underwater small targets.
[0010] Furthermore, the image preprocessing module preprocesses the input image based on an improved dark channel prior algorithm, and the specific steps include:
[0011] Step 1, underwater physical degradation modeling: Based on the wavelength-selective attenuation characteristics of water to light, a light scattering degradation model is established to distinguish the attenuation range of different color channels;
[0012] Step 2, Dark Channel Map Calculation and Optimization: Calculate the minimum channel intensity map in a local area of the input image, and combine it with the suspended object masking technique to eliminate interference from near-field particles and generate a noise-resistant dark channel map;
[0013] Step 3, Intelligent Estimation of Background Light: Based on the noise-resistant dark channel map, the pixels with the brightness in the top 0.1% are selected as candidate points for background light, and depth information is fused to eliminate near-field interference and accurately extract the background light value of the water body.
[0014] Step 4, Transmittance Map Calculation and Filtering: Estimate the initial transmittance map based on the prior rules of the dark channel; perform edge-aware filtering using the green channel as the guide image, and combine the depth prior to constrain the distant transmittance to obtain the corrected transmittance map.
[0015] Step 5, Scene Radiance Restoration: Based on the water background light value obtained in Step 3 and the corrected transmittance map obtained in Step 4, the clear scene radiation is inverted using the light scattering degradation model, and the restored image is subjected to adaptive contrast stretching and edge protection fusion.
[0016] Step 6, Post-processing enhancement: Perform channel equalization and adaptive sharpening to improve the overall clarity of the restored image.
[0017] Furthermore, the backbone network adopts the DarkNet53 architecture, which includes multiple convolutional layers and one fully connected layer, and introduces a residual module in each convolutional layer. The residual module consists of two consecutive 3×3 convolutional layers and a cross-layer skip connection, which adds the input features to the output features of the second convolution element by element through the cross-layer skip connection.
[0018] Furthermore, the context aggregation module uses pooling kernels of different sizes to perform multi-scale pooling operations on the feature maps of each scale output by the feature pyramid network. Then, the output of each pooling branch is upsampled to the spatial size of the corresponding input feature map and concatenated along the channel dimension. Finally, the concatenated features are fused and the number of channels is adjusted through 1x1 convolution to finally output the multi-scale feature map after context aggregation.
[0019] Furthermore, the multi-scale filtering module includes a channel purification module and a spatial purification module. First, the multi-scale feature maps output by the context aggregation module are scale-aligned to obtain multiple input feature maps with consistent scale. Then, the multiple input feature maps with consistent scale are convolved and concatenated along the channel dimension to obtain the concatenated feature map.
[0020] The channel purification module performs average pooling and max pooling operations on the spliced feature map, adds the results of the two pooling operations and normalizes them to generate a set of channel adaptive weights. The channel adaptive weights are used to perform weighted summation on multiple input features with the same scale to output the channel purified feature map.
[0021] The spatial cleanup module performs a Softmax operation on the stitched feature map to generate multiple sets of spatial attention weights, each set of spatial attention weights corresponding to a spatial attention mode. For each set of spatial attention weights, the multiple input feature maps after scale alignment are weighted and summed at different channels and spatial locations. Then, the outputs of all attention modes are summed to generate the spatially cleaned feature map.
[0022] Finally, the feature maps output by the channel purification module and the features output by the space purification module are fused together to form the most comprehensive multi-scale filtering feature output.
[0023] Furthermore, on the Jetson TX2 platform, hardware acceleration is achieved using dual digital signal processing (DSP) cores and a 256-core NVIDIA Pascal GPU.
[0024] The underwater small target detection model is loaded using TensorRT, and inference efficiency is improved by combining INT8 quantization and operator fusion technology. The weights of channel purification and spatial purification of the multi-scale filtering module are calculated in parallel by dual DSP cores. A pre-allocated memory pool strategy is adopted to uniformly manage and reuse feature maps, weights and intermediate results in the inference process, reduce dynamic memory allocation overhead and improve memory access efficiency.
[0025] This invention also protects the construction method of the underwater small target detection system based on dual-cleaning feature fusion, comprising the following steps:
[0026] Step 1: Using the images from the official competition dataset URPC2021 of the Underwater Robot Grasping Competition (URPC) as the dataset, divide the dataset into training set, test set, and validation set according to the ratio of 80%, 10%, and 10%, respectively. The image preprocessing module then preprocesses the images in the acquired dataset.
[0027] Step 2: Construct the underwater small target detection model structure, which is an improved Faster R-CNN network structure. This includes replacing the backbone network of the original Faster R-CNN network structure with the DarkNet53 architecture and introducing a residual module into the DarkNet53 architecture. The context aggregation module and the multi-scale filtering module are connected sequentially between the output of the feature pyramid network and the input of the region proposal network.
[0028] Step 3: Based on the training set, take the preprocessed images as input and the corresponding small target feature images and location information in the images as output to train the underwater small target detection model structure to obtain the trained underwater small target detection model.
[0029] Step 4: Based on the test set, using the preprocessed images as input, apply the trained underwater small target detection model to output the corresponding small target feature images and location information in each image, in order to evaluate and optimize the model.
[0030] Step 5: After optimizing the underwater small target detection model through TensorRT, deploy it on the JetsonTX2 platform to achieve real-time inference.
[0031] Furthermore, during model training, a joint loss function consisting of classification loss and regression loss is used to simultaneously optimize the classification task and the bounding box regression task;
[0032] The classification loss uses the cross-entropy loss function;
[0033] The regression loss adopts an improved Intersection over Union (IoU) Loss function and introduces a scale-sensitive factor and a bounding box non-overlap optimization mechanism;
[0034] The weight coefficients of the two types of losses are dynamically adjusted, and the gradients generated by the two types of losses are normalized to ensure stable convergence during training.
[0035] Furthermore, the evaluation metrics for the underwater small target detection model are average precision (AP), mean average precision (mAP), and frames per second (FPS).
[0036] Furthermore, during deployment, the Jetson TX2 performance monitoring tool was used to perform real-time performance analysis on the model's inference process, obtaining performance indicators including inference time and detection accuracy to ensure that inference time and accuracy meet the requirements of real-time detection; and based on the monitoring results, the model's structural parameters and calculation process were optimized to improve the overall performance of the system.
[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0038] Significantly improved image quality: The suspended object masking technology resists near-field interference, corrects underwater blue-green bias, and preserves the authenticity of suspended particles by dynamically adjusting the tensile strength;
[0039] Improved small target detection accuracy: By effectively aggregating contexts of different ranges through a multi-scale context aggregation mechanism, it helps to understand target boundaries and regions in a blurred background and combat low contrast.
[0040] Feature fusion mechanism optimization: Through the dual-cleanup fusion mechanism, semantic conflicts in multi-scale feature fusion are effectively reduced, the expression of features at different scales is optimized, and the accuracy of target detection is improved;
[0041] Excellent embedded deployment efficiency: On embedded devices, through hardware acceleration and memory optimization, the target detection task is executed efficiently, meeting the needs of real-time detection. Attached Figure Description
[0042] Figure 1 This is an overall architecture diagram of the underwater small target detection model provided by the present invention;
[0043] Figure 2 The context aggregation module structure diagram provided by this invention;
[0044] Figure 3 The structural diagram of the multi-scale filtering module provided by this invention;
[0045] Figure 4 This is a flowchart illustrating the hardware acceleration and embedded deployment process of the embedded platform in this invention.
[0046] Figure 5 A flowchart of the method provided by the present invention. Detailed Implementation
[0047] To make the technical solution of the present invention clearer and more complete, the technical solution of the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.
[0048] like Figures 1 to 4 As shown, the underwater small target detection system based on dual-cleaning feature fusion provided by this invention includes: an image preprocessing module for preprocessing the input underwater image based on an improved dark channel prior algorithm; and an underwater small target detection model, which is an improved Faster algorithm. The R-CNN network comprises a backbone network, a feature pyramid network, a context aggregation module, a multi-scale filtering module, a region proposal network, a ROI pooling layer, and a detection head. The backbone network uses the DarkNet53 architecture to extract primary feature maps from the preprocessed image. The feature pyramid network upsamples the primary feature maps output by the backbone network and fuses them with downsampled feature maps of the corresponding scale to generate multi-scale feature maps. The context aggregation module performs multi-scale pooling operations on the feature maps at each scale output by the feature pyramid network to enhance contextual information representation. The multi-scale filtering module employs a dual-cleaning fusion mechanism of channel cleansing and spatial cleansing to adaptively weight and fuse the multi-scale features output by the context aggregation module. The region proposal network generates candidate regions based on the dual-cleaned fused features. The ROI pooling layer and the detection head classify and regress the bounding boxes of the candidate regions, outputting the final detection results. The underwater small target detection model is deployed on the Jetson TX2 platform with an integrated TensorRT inference engine to accelerate inference.
[0049] Specifically, the image preprocessing module performs preprocessing on the input image based on an improved dark channel prior algorithm. The specific steps include:
[0050] Step 1, underwater physical degradation modeling: Based on the wavelength-selective attenuation characteristics of water to light, a light scattering degradation model is established to clearly distinguish the attenuation degree of red, green and blue channels, where the blue or green channel has the least attenuation and the red channel has the greatest attenuation.
[0051] Step 2, Dark Channel Map Calculation and Optimization: Calculate the minimum intensity value of each pixel in the R, G, and B channels in a local area of the input image to generate an initial dark channel map; combine the suspended object masking technique to eliminate interference from near-field particles and generate a noise-resistant dark channel map;
[0052] Step 3, Intelligent Estimation of Background Light: Based on the noise-resistant dark channel map, the pixels with the brightness in the top 0.1% are selected as candidate points for background light, and depth estimation information is fused to eliminate near-field interference and accurately extract the background light value of the water body.
[0053] Step 4: Transmittance Map Calculation and Filtering. The initial transmittance map is estimated based on the prior rules of the dark channel; edge-aware filtering is performed on the green channel as the guide image to suppress blue-green bias and block effect; and the distant transmittance is constrained by the depth prior to avoid over-enhancement, thereby obtaining the corrected transmittance map.
[0054] Step 5, Scene Radiation Restoration: Based on the water background light value obtained in Step 3 and the corrected transmittance map obtained in Step 4, the clear scene radiation is inverted using the light scattering degradation model. The restored image is then subjected to adaptive contrast stretching, and the original image texture is fused using an edge protection strategy to preserve the authenticity of underwater suspended particles.
[0055] Step 6, Post-processing enhancement: Perform channel equalization to compensate for excessive attenuation of the red channel to improve the visibility of distant red targets; apply adaptive threshold edge enhancement to restore details lost due to scattering and further improve the overall clarity of the restored image.
[0056] Specifically, the backbone network adopts the DarkNet53 architecture, which includes 52 convolutional layers and 1 fully connected layer, and introduces a residual module in each convolutional layer. The residual module consists of two consecutive 3×3 convolutional layers and a cross-layer skip connection, which adds the input features to the output features of the second convolution element by element through the cross-layer skip connection.
[0057] Specifically, such as Figure 2 As shown, the context aggregation module uses three pooling branches with 1×1, 3×3, and 6×6 pooling kernels respectively to perform global average pooling, adaptive average pooling, and adaptive average pooling operations on the feature maps of each scale output by the feature pyramid network. Then, the output of each pooling branch is bilinearly upsampled to the spatial size of the corresponding input feature map and concatenated along the channel dimension. Finally, the concatenated features are fused and the channels are adjusted through 1x1 convolution to output the multi-scale feature map after context aggregation.
[0058] Specifically, the multi-scale filtering module includes a channel purification module and a spatial purification module; the multi-scale feature map output by the context aggregation module is input into the multi-scale filtering module, and the feature maps at each level are adjusted to be consistent with the scale of the target level feature map after upsampling or downsampling;
[0059] like Figure 3As shown, the context aggregation module outputs three input feature maps L1, L2, and L3, with the scales of L1, L2, and L3 increasing sequentially. The figure shows that L1 is upsampled, and L3 is downsampled and adjusted to the same scale as L2, resulting in three input feature maps R1, R2, and R3 with consistent scales. In practice, feature maps at each level can be adjusted to any target level as needed. scale and ;
[0060] Adjust the three to the target level. The scaled input feature maps are convolutionally processed and then concatenated along the channel dimension to obtain the concatenated feature maps. ;
[0061] The channel purification module performs average pooling and max pooling operations on the concatenated feature maps, then adds the results of the two pooling operations element by element, and normalizes them using the Sigmoid activation function to generate adaptive channel weights.
[0062] ,
[0063] Using the generated channel adaptive weights, the three adjusted input feature maps are weighted and fused, and the positions are... The output feature vector is represented as:
[0064] ,
[0065] In the formula, For the first The input feature map is adjusted to the _th _ ... After the layer in the position The eigenvalues at, where ; For the channel purification module at the location Output feature vector at; , , The channel adaptive weights for each input feature map are adjusted. This represents the Sigmoid operation. This indicates that average pooling is performed on the concatenated feature maps. This indicates that max pooling is performed on the concatenated feature map;
[0066] The spatial cleanup module uses a Softmax operation to generate relative weights for all positions relative to the channels in the stitched feature map; then, it performs weighted fusion on the input feature map, considering the positions... The output feature is represented as follows:
[0067] ,
[0068] In the formula, For the space purification module in location The output vector at that point; The channel index representing the spatial attention weights and This corresponds to three different spatial attention modes; , , They represent the first Attention patterns in location Spatial attention weights for each input feature map after adjustment; The channel index of the input feature map; For the first The input feature map is adjusted to the _th _ ... After the layer, in the first passage, location The eigenvalues at, where ; Used to normalize the feature mapping along the channel direction to obtain the relative weights of different channels at the same location;
[0069] Total output of the multi-scale filtering module Represented as:
[0070] ,
[0071] in , These are the total output characteristics of the channel purification module and the total output characteristics of the space purification module, respectively.
[0072] By employing a dual-cleanup fusion mechanism, semantic conflicts in multi-scale feature fusion are effectively reduced, the expression of features at different scales is optimized, and the accuracy of target detection is improved.
[0073] Specifically, such as Figure 4 As shown, on the Jetson TX2 platform, hardware acceleration is achieved using dual DSP cores and a 256-core NVIDIA Pascal GPU, reducing power consumption and improving computing efficiency.
[0074] An underwater small target detection model was loaded using TensorRT, and the model weights were quantized using TensorRT's INT8 quantization function. Simultaneously, operator fusion technology was used to optimize the computation flow, significantly reducing memory usage and computational overhead. During model operation, parallel computation of the multi-scale filtering module was performed using dual DSP cores to accelerate weight calculations in the channel and spatial purification modules. Furthermore, a pre-allocated memory pool strategy was adopted to uniformly manage and reuse feature maps, weights, and intermediate results during inference, reducing dynamic memory allocation overhead and improving memory access efficiency.
[0075] like Figure 5 As shown, the method for constructing an underwater small target detection system based on dual-cleaning feature fusion provided by the present invention includes the following steps:
[0076] Step 1: Using the images from the official competition dataset URPC2021 of the Underwater Robot Grasping Competition (URPC) as the dataset, divide the dataset into training set, test set, and validation set according to the ratio of 80%, 10%, and 10%, respectively. The image preprocessing module then preprocesses the images in the acquired dataset.
[0077] Step 2: Construct the underwater small target detection model structure, which is an improved Faster R-CNN network structure. This includes replacing the backbone network of the original Faster R-CNN network structure with the VGG16 network and the DarkNet53 architecture, and introducing a residual module into the DarkNet53 architecture. A context aggregation module and a multi-scale filtering module are connected sequentially between the output of the feature pyramid network and the input of the region proposal network.
[0078] Step 3: Based on the training set, take the preprocessed images as input and the corresponding small target feature images and location information in the images as output to train the underwater small target detection model structure to obtain the trained underwater small target detection model.
[0079] Step 4: Based on the test set, using the preprocessed images as input, apply the trained underwater small target detection model to output the corresponding small target feature images and location information in each image, in order to evaluate and optimize the model.
[0080] Step 5: After optimizing the underwater small target detection model through the TensorRT inference engine, deploy it on the Jetson TX2 platform to achieve real-time inference.
[0081] Specifically, the model structure was trained in a Python 3.8 environment, based on the deep learning frameworks PyTorch 1.7.0 and torchvision 0.7.0; the optimizer was AdamW, and the training period was 120 epochs; during training, a joint loss function consisting of classification loss and regression loss was used to simultaneously optimize the classification task and the bounding box regression task; after training, average precision (AP), mean average precision (mAP), and frames per second (FPS) were used as evaluation metrics to evaluate the optimized model.
[0082] Specifically, during embedded deployment, the trained underwater small target detection model is first exported to ONNX format using the PyTorch framework, and then loaded and converted by TensorRT to generate a TensorRT inference engine suitable for the Jetson TX2 platform.
[0083] The underwater small target detection model optimized by TensorRT is deployed on the Jetson TX2 platform. The TensorRT inference engine is called through Python or C++ to perform real-time target detection and output detection boxes and classification results in real time.
[0084] Specifically, during deployment, Jetson TX2's performance monitoring tool is used for real-time performance analysis to ensure that inference time and accuracy meet the requirements of real-time detection; and based on the monitoring results, the model and calculation process are further optimized to improve the overall performance of the system.
[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made by those skilled in the art within the scope of the technology disclosed in this invention, based on the technical solution and concept of the present invention, should be included within the protection scope of this invention. Therefore, the protection scope of this invention should be determined by the scope of the claims.
Claims
1. An underwater small target detection system based on dual-purification feature fusion, characterized in that, include: The image preprocessing module is used to preprocess the input underwater images based on an improved dark channel prior algorithm. The underwater small target detection model is an improved Faster R-CNN network, which includes a backbone network, a feature pyramid network, a context aggregation module, a multi-scale filtering module, a region proposal network, a ROI pooling layer, and a detection head. The backbone network adopts the DarkNet53 architecture to extract primary feature maps from the preprocessed image. The feature pyramid network is used to upsample the primary feature map output by the backbone network and fuse it with the downsampled feature map at the corresponding scale to generate a multi-scale feature map. The context aggregation module performs multi-scale pooling operations on the feature maps at each scale output by the feature pyramid network to enhance the representation of contextual information. The multi-scale filtering module adopts a dual-cleaning fusion mechanism of channel cleanup and spatial cleanup to adaptively weight and fuse the multi-scale features output by the context aggregation module. The region proposal network generates candidate regions based on the dual-cleaned fused features, and completes classification and bounding box regression through the ROI pooling layer and the detection head to output the final detection results. The underwater small target detection model is optimized and deployed on the Jetson TX2 platform through TensorRT to realize real-time inference of underwater small targets.
2. The underwater small target detection system based on dual purification feature fusion according to claim 1, characterized in that, The image preprocessing module preprocesses the input image based on an improved dark channel prior algorithm. The specific steps include: Step 1, underwater physical degradation modeling: Based on the wavelength-selective attenuation characteristics of water to light, a light scattering degradation model is established to distinguish the attenuation range of different color channels; Step 2, Dark Channel Map Calculation and Optimization: Calculate the minimum channel intensity map in a local area of the input image, and combine it with the suspended object masking technique to eliminate interference from near-field particles and generate a noise-resistant dark channel map; Step 3, Intelligent Estimation of Background Light: Based on the noise-resistant dark channel map, the pixels with the brightness in the top 0.1% are selected as candidate points for background light, and depth information is fused to eliminate near-field interference and accurately extract the background light value of the water body. Step 4, Transmittance Map Calculation and Filtering: Estimate the initial transmittance map based on the prior rules of the dark channel; perform edge-aware filtering using the green channel as the guide image, and combine the depth prior to constrain the distant transmittance to obtain the corrected transmittance map. Step 5, Scene Radiance Restoration: Based on the water background light value obtained in Step 3 and the corrected transmittance map obtained in Step 4, the clear scene radiation is inverted using the light scattering degradation model, and the restored image is subjected to adaptive contrast stretching and edge protection fusion. Step 6, Post-processing enhancement: Perform channel equalization and adaptive sharpening to improve the overall clarity of the restored image.
3. The underwater small target detection system based on dual purification feature fusion according to claim 1, characterized in that, The backbone network adopts the DarkNet53 architecture, which includes multiple convolutional layers and one fully connected layer, and introduces a residual module in each convolutional layer. The residual module consists of two consecutive 3×3 convolutional layers and a cross-layer skip connection, which adds the input features to the output features of the second convolution element by element through the cross-layer skip connection.
4. The underwater small target detection system based on dual purification feature fusion according to claim 1, characterized in that, The context aggregation module uses pooling kernels of different sizes to perform multi-scale pooling operations on the feature maps of each scale output by the feature pyramid network. Then, the output of each pooling branch is upsampled to the spatial size of the corresponding input feature map and concatenated along the channel dimension. Finally, the concatenated features are fused and the number of channels is adjusted through 1x1 convolution to finally output the multi-scale feature map after context aggregation.
5. The underwater small target detection system based on dual purification feature fusion according to claim 4, characterized in that, The multi-scale filtering module includes a channel purification module and a spatial purification module. First, the multi-scale feature maps output by the context aggregation module are scale-aligned to obtain multiple input feature maps with consistent scale. Multiple input feature maps of the same scale are convolved and concatenated along the channel dimension to obtain the concatenated feature map; The channel purification module performs average pooling and max pooling operations on the spliced feature map, adds the results of the two pooling operations and normalizes them to generate a set of channel adaptive weights. The channel adaptive weights are used to perform weighted summation on multiple input features with the same scale to output the channel purified feature map. The spatial purification module performs a Softmax operation on the spliced feature map to generate multiple sets of spatial attention weights, each set of spatial attention weights corresponding to a spatial attention mode. For each set of spatial attention weights, the multiple input feature maps after scale alignment are weighted and summed at different channels and spatial locations; then the outputs of all attention modes are summed to generate spatially cleaned feature maps. Finally, the feature maps output by the channel purification module and the features output by the space purification module are fused together to form the final multi-scale filtering feature output.
6. The underwater small target detection system based on dual purification feature fusion according to claim 1, characterized in that, On the Jetson TX2 platform, hardware acceleration is achieved using dual digital signal processing cores and a 256-core NVIDIA Pascal GPU. We utilize TensorRT to load an underwater small target detection model and combine INT8 quantization and operator fusion techniques to improve inference efficiency. The weights of channel purification and spatial purification in the multi-scale filtering module are calculated in parallel using dual DSP cores. It also adopts a pre-allocated memory pool strategy to uniformly manage and reuse feature maps, weights, and intermediate results during the inference process.
7. The method for constructing an underwater small target detection system based on dual-cleaning feature fusion according to any one of claims 1 to 6, characterized in that, Includes the following steps: Step 1: Using the images captured by the underwater robot as the dataset, divide the dataset into training set, test set, and validation set according to the ratio of 80%, 10%, and 10%, respectively. The image preprocessing module then preprocesses the images in the acquired dataset. Step 2: Build the underwater small target detection model structure, which is an improved Faster R-CNN network structure. This includes replacing the backbone network of the original Faster R-CNN network structure with the VGG16 network and the DarkNet53 architecture, and introducing residual modules into the DarkNet53 architecture. A context aggregation module and a multi-scale filtering module are sequentially connected between the output of the feature pyramid network and the input of the region proposal network. Step 3: Based on the training set, take the preprocessed images as input and the corresponding small target feature images and location information in the images as output to train the underwater small target detection model structure to obtain the trained underwater small target detection model. Step 4: Based on the test set, using the preprocessed images as input, apply the trained underwater small target detection model to output the corresponding small target feature images and location information in each image, in order to evaluate and optimize the model. Step 5: After optimizing the underwater small target detection model through TensorRT, deploy it on the Jetson TX2 platform to achieve real-time inference.
8. The method for constructing an underwater small target detection system based on dual purification feature fusion according to claim 7, characterized in that, During model training, a joint loss function consisting of classification loss and regression loss is used to simultaneously optimize the classification task and the bounding box regression task. The classification loss uses the cross-entropy loss function; The regression loss adopts an improved cross-union ratio loss function and introduces a scale-sensitive factor and a bounding box non-overlapping optimization mechanism; The weight coefficients of the two types of losses are dynamically adjusted, and the gradients generated by the two types of losses are normalized.
9. The method for constructing an underwater small target detection system based on dual purification feature fusion according to claim 7, characterized in that, The evaluation metrics for underwater small target detection models are average accuracy, mean average accuracy, and frames processed per second.
10. The method for constructing an underwater small target detection system based on dual-cleaning feature fusion according to claim 7, characterized in that, During deployment, the Jetson TX2 performance monitoring tool was used to perform real-time performance analysis on the model's inference process and obtain performance metrics including inference time and detection accuracy. Based on the monitoring results, the structural parameters and calculation process of the model are optimized to improve the overall performance of the system.