A biomimetic vision-based image enhancement device, underwater target detection system and method

Abstract

This invention discloses an image enhancement device, underwater target detection system, and method based on biomimetic vision. It simulates the visual perception mechanism of higher visual organisms to mimic the mechanisms by which these organisms extract key information from complex visual scenes. Furthermore, it mimics the characteristic of these organisms to give different levels of attention to targets based on their importance and location when processing different visual information. This improves upon existing underwater target detection models' difficulties in detecting small underwater targets in complex environments and when there is significant similarity between the target and the background. This invention enhances target detection accuracy without significantly increasing computational complexity, making it more suitable for real-time applications in underwater environments.

Description

Technical Field

[0001] This invention relates to the field of underwater target detection technology, specifically to an image enhancement device, underwater target detection system, and method based on biomimetic vision. Background Technology

[0002] Underwater target detection is a crucial application of computer vision in underwater environments, widely used in marine resource surveying, ecological environment monitoring, underwater robot navigation, and marine rescue. However, the complexity of the underwater environment presents numerous challenges to target detection, including uneven lighting, color degradation, and low contrast. Furthermore, underwater targets are typically small and easily obscured by complex backgrounds, making feature extraction and target recognition even more difficult. Therefore, designing an efficient target detection feature extraction module for complex underwater environments is a key issue for improving detection performance.

[0003] In recent years, deep learning technologies, especially object detection algorithms based on convolutional neural networks (CNNs), such as YOLO (You Only Look Once), Faster R-CNN, and SSD (Single Shot MultiBox Detector), have made some progress in underwater object detection tasks. However, these networks are mainly designed for natural scenes, and their direct application to underwater environments still presents the following problems: underwater light attenuation and scattering reduce image contrast, making it difficult for traditional convolutional layers to extract effective features, thus affecting detection accuracy; underwater targets are usually small, such as fish, corals, and plankton, while traditional detection networks mainly optimize for larger targets and lack the ability to effectively extract features from small targets, leading to small targets being easily overlooked; underwater scenes contain a large number of background elements similar to targets, such as seaweed, rocks, and plankton, which are similar to targets in color, shape, etc., easily leading to false detections; existing detection networks are usually computationally expensive and not suitable for edge devices with limited computing power (such as underwater drones and embedded platforms). How to optimize computational efficiency while maintaining detection accuracy has become an important research direction.

[0004] To address the challenges of underwater target detection, existing research mainly explores solutions through the following methods: (1) Underwater image enhancement. Due to uneven illumination, color decay, and low contrast in underwater environments, many studies focus on image preprocessing and enhancement techniques. For example, physical model-based methods (such as the DCP dehazing algorithm based on the light propagation model) attempt to restore true colors, but their generalization to different aquatic environments is weak; data-driven methods (such as GAN-based image enhancement) improve visual quality by learning end-to-end mappings, but may introduce artifacts, affecting the stability of detection. In addition, brightness correction methods that incorporate Retinex theory can improve local contrast, but are still difficult to adapt to complex underwater optical conditions. (2) Feature extraction. Many studies use high-resolution feature maps to preserve detailed information or employ multi-branch convolutions (such as Inception-like structures) to extract local features at different scales simultaneously. However, these methods usually increase computational complexity and may lead to information redundancy, affecting detection efficiency. Using architectures with strong global perception capabilities, such as Transformer (e.g., Swin-Transformer), can capture long-distance dependencies, but its computational complexity is high, making it difficult to run in real time on edge devices. (3) Network structure optimization. In recent years, researchers have proposed a series of schemes to optimize network structures in order to improve the robustness of underwater target detection. For example, optimization methods based on attention mechanisms (e.g., SE, CBAM, ECA) can enhance attention to key regions and improve the ability to identify targets. However, traditional attention mechanisms are mainly designed based on natural scenes and may not be able to separate targets from the background well in underwater environments. In addition, researchers have also tried to introduce improved backbone networks (e.g., EfficientNet, MobileNetV3) to improve the feature extraction capability of the model while reducing computational costs, but these networks still have difficulty solving the problem of small target detection. Summary of the Invention

[0005] To address the difficulties existing underwater target detection models face in detecting complex environments, small underwater targets, and targets with similar backgrounds, the present invention aims to provide an image enhancement device, underwater target detection system, and method based on biomimetic vision to improve underwater target detection capabilities.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A biomimetic vision-based image enhancement device includes a region perception module, a multi-scale feature aggregation module, a convolutional block attention mechanism module, and a contrast adaptive normalization module.

[0008] The region perception module mimics the functions of the central and peripheral retina in advanced biological vision systems. The central retina is responsible for perceiving fine target details, while the peripheral retina is used to perceive the overall scene and quickly locate targets.

[0009] The multi-scale feature aggregation module simulates the ability of the biological visual system to process information at different scales and depths, helping to extract and aggregate features at multiple scales, thereby comprehensively capturing image details;

[0010] The convolutional block attention mechanism module simulates the attention mechanism of advanced biological visual systems, enabling it to dynamically focus on important features, suppress irrelevant information, and enhance the expressive power of key features through channel and spatial attention.

[0011] The contrast adaptive normalization module simulates the mechanism of adaptive contrast adjustment in biological visual systems, automatically adjusting image brightness and contrast according to different lighting conditions to enhance the visibility of important features.

[0012] The region perception module enhances the input feature map to obtain an enhanced feature map; the multi-scale feature aggregation module processes the enhanced feature map and fuses local and global features through different receptive field features to obtain fused features; the convolutional block attention mechanism module further filters the fused features to select important feature information; the contrast adaptive normalization module enhances the selected important feature information to obtain the final enhanced features.

[0013] The region perception module processes the image as follows:

[0014] Assuming the input feature map The size is ,in It's about height. It's the width. It refers to the number of channels; firstly, a 3×3 convolution is used to simulate the central retina's focusing on the target details, extracting local detail features, i.e.

[0015] ;

[0016] in, This represents a 3×3 convolution. Indicates background features, Indicates the input feature map;

[0017] Next, a 5×5 convolution is used to simulate the peripheral retina's perception of the global background, extracting background features over a larger area:

[0018]

[0019] in, This represents a 5×5 convolution. Indicates background features, The input feature map is represented; an attention map is generated using the Sigmoid activation function. :

[0020] ;

[0021] in, This represents the activation function. For local features, Background features;

[0022] Input feature map and attention map Element-wise multiplication is performed to enhance the features of the target region and suppress the features of the background region, resulting in an attention-weighted feature map:

[0023]

[0024] in, Feature maps weighted for attention, For the input feature map, This is an attention graph.

[0025] The region-aware module uses depthwise separable convolutions to replace traditional convolutions, and the 5×5 convolutions are replaced by 3×3 dilated convolutions with a dilation rate of 2.

[0026] The processing of the multi-scale feature aggregation module is as follows:

[0027] First, three depthwise separable convolutions with dilation rates of 1, 2, and 3 are used to extract feature information from different receptive fields:

[0028] ;

[0029] ;

[0030] ;

[0031] in, These represent the extracted feature information at three different scales. This represents depthwise separable convolution. Feature maps weighted for attention, Indicates the expansion rate;

[0032] Next, global average pooling is performed to obtain global information from the input feature map;

[0033] ;

[0034] in, Represents global average pooling; This represents global information about the input feature map. Attention-weighted feature maps;

[0035] Then, the pooling result is transformed into a one-dimensional feature representation, and weights at three scales are computed through a fully connected layer:

[0036]

[0037] in, Represents the weights of the three scales. Indicates a fully connected layer. This represents global information about the input feature map. express Activation function;

[0038] Using the calculated weights Weighted fusion of multi-scale features:

[0039]

[0040] in, This is the weighted feature map; For expansion rate The feature weights extracted at that time, Indicates the expansion rate Feature information extracted in time;

[0041] The weighted feature maps are then further fused using a 1×1 convolution, and the fused feature maps are residually connected to the input feature maps to avoid information loss.

[0042]

[0043] in, To fuse feature maps, This is the weighted feature map. Feature maps weighted for attention, This represents a 1×1 convolution.

[0044] The convolutional block attention mechanism module includes a channel attention unit and a spatial attention unit;

[0045] The main objective of the channel attention unit is to weight each channel of the feature map according to the importance of each channel, with the weight of each channel reflecting the importance of that channel in the current task.

[0046] The goal of the spatial attention unit is to dynamically focus on important regions in the feature map by introducing an attention mechanism in the spatial dimension.

[0047] The specific implementation steps of the channel attention unit are as follows:

[0048] Global pooling: First, the input feature map is processed by global average pooling and global max pooling to generate a global description; global average pooling calculates the average value of each channel to capture the global information of the image; while global max pooling calculates the maximum value of each channel to capture the most prominent features.

[0049] Fully connected layer: Next, the two pooling results are input into a two-layer fully connected network; the first layer reduces the dimensionality of the number of channels, and the second layer restores the number of channels. Each channel is assigned a weight representing its importance after passing through this network.

[0050] Activation function: Finally, a channel-level attention weight coefficient is obtained by passing the sigmoid activation function; this weight coefficient is then multiplied channel by channel with the input feature map to obtain the adjusted feature map;

[0051]

[0052] in, Indicates channel attention. , These represent the weight matrices of the two fully connected layers, This represents the sigmoid activation function. Represents the ReLU activation function. Represents global average pooling. Represents global max pooling. Feature maps are fused.

[0053] The specific implementation of the spatial attention module is as follows:

[0054] Channel operations on feature maps: By performing operations on the input feature map along the channel dimension, the importance of each position is calculated; average pooling is used to obtain global information along the channel dimension, and max pooling is used to obtain salient features along the channel dimension.

[0055] Convolution operation: The feature maps of the two channels are concatenated, and a 7×7 convolution is used to extract spatial relationships over a large range, generating a spatial attention map. The output is an attention map with the same spatial dimension as the input feature map, representing the importance of each spatial location.

[0056] Activation function: Finally, the values ​​of the spatial attention map are mapped to the [0,1] interval through the Sigmoid activation function, and pixel-wise multiplication is performed with the spatial dimension of the input feature map to enhance the features of important regions and suppress unimportant regions;

[0057]

[0058] in, Indicates channel attention, Indicating spatial attention, This represents the attention of the i-th channel. Indicates the number of channels. This represents the sigmoid activation function. This represents convolution.

[0059] The contrast adaptive normalization module obtains the brightness information of each channel through global mean pooling, and then adaptively adjusts the image based on this brightness information. Specifically, the value of each pixel in the image is adjusted differently from the global mean of its channel. By scaling the scale and bias parameters, the adjusted image brightness adapts to different input scenarios.

[0060] ;

[0061] in, This indicates the final output. Indicates spatial attention. Indicates global average pooling. Indicates the scaling factor. This represents the bias parameter.

[0062] An underwater target detection system based on biomimetic vision includes a target detection model, which comprises a backbone network, a neck network, and a head network, wherein the image enhancement device described above is incorporated into the neck network.

[0063] A biomimetic vision-based underwater target detection method is proposed, which is implemented using an underwater target detection system based on biomimetic vision as described above. The method specifically comprises:

[0064] The raw underwater image is input into the target detection model. The backbone network of the target detection model extracts the basic features of the shallow and deep layers in the raw underwater image, and then sends them to the neck network for feature fusion. The image features are enhanced by an image enhancement device, and finally the enhanced features are input into the head network for classification and localization to perform target detection.

[0065] By adopting the above scheme, this invention simulates the visual perception mechanism of advanced visual organisms, mimicking their extraction of key information from complex visual scenes. It also imitates the characteristic that advanced visual animals often give different levels of attention based on the importance and location of targets when processing different visual information. This improves upon existing underwater target detection models' difficulties in detecting complex environments, small underwater targets, and targets with similar backgrounds. This invention enhances target detection accuracy without significantly increasing computational complexity, making it more suitable for real-time applications in underwater environments. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the implementation of the system of the present invention;

[0067] Figure 2 This is a structural diagram of an underwater target detection system;

[0068] Figure 3 A flowchart for underwater target detection;

[0069] Figure 4 A comparison chart of experimental results indicators;

[0070] Figure 5 The experimental results are shown in comparison (left: YOLOv11n, right: YOLOv11n+BVM).

[0071] Figure 6 A comparison of output heatmaps before and after adding BVM. Detailed Implementation

[0072] like Figure 1 As shown, this invention discloses an image enhancement device based on biomimetic vision, which includes a region perception module, a multi-scale feature aggregation module, a convolutional block attention mechanism module, and a contrast adaptive normalization module. The region perception module generates an enhanced feature map from the input feature map and sends it to the multi-scale feature aggregation module to further fuse local and global features through different receptive field features. The fused features are then sent to the convolutional block attention mechanism module to further filter important feature information. Finally, the features are sent to the contrast adaptive normalization module for final feature enhancement, assisting the detector in obtaining clearer target features.

[0073] The system comprises the following modules: **Regional Perception Module (RPM):** This module mimics the functions of the central and peripheral retina in advanced biological vision systems. The central retina is responsible for perceiving fine target details, while the peripheral retina is used to perceive the overall scene and quickly locate targets. **Multi-Scale Feature Aggregation (MSFA):** This module simulates the ability of biological vision systems to process information at different scales and depths, helping to extract and aggregate features at multiple scales to comprehensively capture image details. **Convolutional Block Attention Module (CBAM):** This module simulates the attention mechanism of advanced biological vision systems, dynamically focusing on important features, suppressing irrelevant information, and enhancing the expressive power of key features through channel and spatial attention. **Contrast Adaptive Normalization Module (CANM):** This module simulates the adaptive contrast adjustment mechanism in biological vision systems, automatically adjusting image brightness and contrast according to different lighting conditions to enhance the visibility of important features.

[0074] The low contrast and complex backgrounds of underwater images necessitate that visual systems dynamically adjust focus within the image to enhance detail perception in low-contrast areas. The underwater environment also presents numerous small targets with minimal difference between them and the background; therefore, focusing on the features of the target region is crucial.

[0075] In advanced biological visual systems, the fovea is responsible for high-resolution, fine visual perception, primarily capturing detailed information about targets, while the peripheral retina is used to perceive the overall scene, enabling rapid target localization but with lower resolution. This division of labor allows advanced visual animals to efficiently balance local detail and global perception during visual cognition, improving the accuracy of target detection. Region perception modules borrow this mechanism to help the network automatically focus on key areas within small targets or complex backgrounds, enhancing the detection capability of small underwater targets.

[0076] Specifically, the region-aware module processes images as follows:

[0077] Assuming the input feature map The size is ,in It's about height. It's the width. It refers to the number of channels. First, a 3×3 convolution (central retinal detail focusing) is used to extract local features. :

[0078]

[0079] in, This represents a 3×3 convolution. Indicates background features, This represents the input feature map.

[0080] Next, a 5×5 convolution (peripheral retinal background enhancement) is used to extract background features over a larger area. :

[0081]

[0082] in, This represents a 5×5 convolution. Indicates background features, This represents the input feature map.

[0083] Attention maps are generated using the Sigmoid activation function. :

[0084]

[0085] in, This represents the activation function. For local features, Background features.

[0086] The input feature map and the attention map are multiplied element-wise to enhance the features of the target region and suppress the features of the background region, resulting in an attention-weighted feature map:

[0087]

[0088] in, Feature maps weighted for attention, For the input feature map, This is an attention graph.

[0089] Meanwhile, to make the module more lightweight, depthwise separable convolutions were used to replace traditional convolutions, and 3×3 dilated convolutions with an expansion rate of 2 were used to replace 5×5 convolutions, thereby reducing computational costs.

[0090] Underwater images are often affected by light attenuation, causing some areas to appear dark and blurry, making it difficult to discern image details. Furthermore, underwater objects exhibit significant scale variations. Therefore, multi-scale feature aggregation can help extract and aggregate information at different scales, capture diverse details, and reduce background noise interference. Biological vision systems are highly complex and optimized perceptual systems, a key characteristic of which is their ability to process information at different levels and scales. The multi-scale feature aggregation module is designed based on this, incorporating different dilated convolutions and global pooling to simulate the multi-scale information processing capabilities of biological vision systems.

[0091] The multi-scale feature aggregation module first uses three depthwise separable convolutions with dilation rates of 1, 2, and 3 to extract multi-scale information. The input is the attention-weighted feature map output by the region awareness module. :

[0092]

[0093]

[0094]

[0095] in These represent the extracted feature information at three different scales. This represents depthwise separable convolution. Indicates the expansion rate.

[0096] Next, a global average pooling operation is performed to obtain global information about the input feature map.

[0097]

[0098] in, Represents global average pooling; This represents global information about the input feature map. Feature maps weighted for attention.

[0099] The pooling result is then transformed into a one-dimensional feature representation, and mapped to a three-dimensional weight space through a fully connected layer to compute weights at three scales:

[0100]

[0101] in, Represents the weights of the three scales. Indicates a fully connected layer. This represents global information about the input feature map. express Activation function.

[0102] Using the calculated weights Weighted fusion of multi-scale features:

[0103] .

[0104] in, This is the weighted feature map; For expansion rate The feature weights extracted at that time, Indicates the expansion rate The feature information extracted in time.

[0105] Weighted feature map Further feature fusion is performed using a 1×1 convolution, and the fused feature map is residually connected to the input feature map to avoid information loss.

[0106] .

[0107] in, For the final fusion feature, This is the weighted feature map. Feature maps weighted for attention, This represents a 1×1 convolution.

[0108] Underwater targets are often obscured by complex backgrounds, especially since background elements such as algae and plankton in images can be very similar in color and shape to the target. Therefore, to more effectively extract and identify underwater targets, a convolutional block attention mechanism module is introduced. This module enhances the network's focus on important features by incorporating channel attention and spatial attention, while suppressing irrelevant or redundant features, thereby improving the expressive power of the feature map and enhancing the accuracy of the underwater target detection network.

[0109] The convolutional block attention mechanism module includes channel attention units and spatial attention units.

[0110] The main objective of the channel attention unit is to weight each channel of the feature map according to its importance. The weight of each channel reflects its importance in the current task. The specific implementation steps are as follows:

[0111] Global Pooling: First, the input feature map is processed by global average pooling and global max pooling to generate a global description. Global average pooling calculates the average value of each channel, capturing global information of the image; while global max pooling calculates the maximum value of each channel, capturing the most prominent features.

[0112] Fully Connected Layer: Next, the two pooling results are fed into a two-layer fully connected network. The first layer reduces the dimensionality of the channels, and the second layer restores the channel count. Each channel receives a weight representing its importance after passing through this network.

[0113] Activation Function: Finally, a channel-level attention weight coefficient is obtained by applying the sigmoid activation function. This weight coefficient is then multiplied channel-by-channel with the input feature map to obtain the adjusted feature map.

[0114] The purpose of channel attention is to enable the network to focus on more important channels and suppress less important channels, thereby improving the expressive power of features.

[0115]

[0116] in, Indicates channel attention. , These represent the weight matrices of the two fully connected layers, This represents the sigmoid activation function. Represents the ReLU activation function. Represents global average pooling. Represents global max pooling. Feature maps are fused.

[0117] The goal of the spatial attention unit is to dynamically focus on important regions in the feature map by introducing an attention mechanism in the spatial dimension. The specific implementation is as follows:

[0118] Channel operations on feature maps: By performing operations on the input feature map along the channel dimension, the importance of each position is calculated. Average pooling is used along the channel dimension to obtain global information, and max pooling is used to obtain salient features.

[0119] Convolution operation: The feature maps of the two channels are concatenated, and a 7×7 convolution is used to extract spatial relationships over a large range, generating a spatial attention map. The output is an attention map with the same spatial dimension as the input feature map, representing the importance of each spatial location.

[0120] Activation function: Finally, the values ​​of the spatial attention map are mapped to the [0,1] interval through the Sigmoid activation function, and pixel-wise multiplication is performed with the spatial dimension of the input feature map to enhance the features of important regions and suppress unimportant regions.

[0121]

[0122] in, Indicates channel attention, Indicates spatial attention. This represents the attention of the i-th channel. Indicates the number of channels. This represents the sigmoid activation function. This represents convolution.

[0123] The contrast adaptive normalization module is designed to enhance the brightness and contrast characteristics of different areas in an image by adaptively adjusting the image's contrast. This is particularly beneficial when processing underwater images, helping to reduce the impact of uneven lighting, blurring, and color distortion. Its primary goal is to enhance target visibility through adaptive adjustment of image brightness.

[0124] The contrast-adaptive normalization module obtains the brightness information of each channel through global mean pooling, and then adaptively adjusts the image based on this brightness information. Specifically, each pixel value in the image is adjusted differentially with the global mean of its corresponding channel. Through scaling and bias parameters, the adjusted image brightness adapts to different input scenarios. By using adaptive scaling and bias parameters, this module can enhance image contrast, improve the representation of details in the image, and make the image brightness adaptable to different scene requirements in different areas, thereby improving the accuracy of object detection or other tasks.

[0125]

[0126] in, This indicates the final output. The spatial attention feature map represents the input. Indicates global average pooling. Indicates the scaling factor. This represents the bias parameter.

[0127] In summary, this invention simulates the visual perception mechanisms of advanced visual organisms, mimicking their extraction of key information from complex visual scenes. It also imitates their tendency to prioritize different visual information based on the importance and location of the target, thereby improving existing underwater target detection models' difficulties in complex environments, small underwater targets, and situations where the target and background are similar. This invention enhances target detection accuracy without significantly increasing computational complexity, making it more suitable for real-time applications in underwater environments.

[0128] The underwater target detection system includes a target detection model, which comprises a backbone network, a neck network, and a head network. When applying the image enhancement device of this invention to underwater target detection, the image enhancement device is simply added to the neck network. For example, as... Figure 2 As shown, the image enhancement device (BVM) is added to the neck network, positioned before the head network. Other integration methods exist, such as integrating it into the network architecture (e.g., C3k2) or other parts of the network. The specific insertion method is flexible and can be adjusted according to the actual situation. Figure 2 In this context, Input represents the input, Conv represents convolution, C3k2 represents a lightweight feature extraction structure, C2PSA represents a lightweight feature enhancement module that integrates spatial attention mechanism, SPPF represents an efficient feature enhancement module, Concat represents feature map concatenation, Upsample represents upsampling, and Detect represents a detection head.

[0129] like Figure 3 As shown, when performing underwater target detection, the original underwater image is input into the target detection model. The backbone network of the target detection model extracts the basic features of the shallow and deep layers in the original underwater image, and then sends them to the neck network for feature fusion. The image features are enhanced by an image enhancement device, and finally the enhanced features are input into the head network for classification and localization to perform target detection.

[0130] To verify the effectiveness of the proposed module, experiments were conducted on the publicly available underwater dataset DUO. The dataset contains 7782 precisely labeled images, of which 6671 were used for training and 1111 for testing. The dataset contains four categories and 74515 objects, with 7887 (10.6%) of sea cucumbers, 50156 (67.3%) of sea urchins, 1924 (2.6%) of scallops, and 14548 (19.5%) of starfish. The DUO dataset contains a diverse range of underwater images and corresponding annotations, covering various underwater scenes including coral reefs, seabed sediments, and shipwrecks, and including objects of different shapes, sizes, and colors. The images in the dataset exhibit typical underwater image characteristics such as color cast, low contrast, uneven lighting, blurriness, and high noise, posing a challenge to the accurate detection of different aquaculture organisms and largely reflecting the problems faced in detecting targets in real marine environments.

[0131] This invention uses the advanced target detector YOLOv11n as a baseline. To address the challenges of complex underwater environments, numerous small targets, and object occlusion, a designed image enhancement device is added before the small target detection head output of the YOLOv11n detector. This enhances its underwater target detection capability while maintaining a sufficiently lightweight model. Regarding experimental parameters...

[0132] This invention sets the number of training epochs to 300, the batch size to 16, uses Adam as the optimizer, the initial learning rate of the model is 0.001, and the weight decay coefficient of the optimizer is 0.0005. For experimental hardware, this invention uses one NVIDIA GeForce RTX 4090 GPU for training, PyTorch version 1.11.0, and Python version 3.8.

[0133] In addition, to further compare the superiority of the designed module, the YOLOv11n version with image enhancement device was compared with the detectors of various lightweight YOLO series versions.

[0134] The key to this invention lies in its dynamic feature focusing mechanism. By flexibly adjusting attention weights, it can adaptively focus on key regions according to the varying complexity of samples, thereby improving the accuracy of underwater target detection. The module effectively enhances its ability to recognize complex backgrounds and small targets through multi-level feature fusion, including region perception, multi-scale feature aggregation, and convolutional block attention mechanisms. Furthermore, the bionic vision module reduces fluctuations during training and ensures the stability of feature learning through dynamic weight generation and adaptive smoothing optimization strategies. Figure 4-6 As shown, the experimental results demonstrate that the addition of an advanced detector to the bionic vision module effectively improves the detector's performance without incurring excessive computational burden. Figure 4 In this context, mAP@0.5 refers to the average AP for each class when the IoU (Intersection over Union) threshold is 0.5. mAP@0.5:0.95 refers to calculating the average AP for each class at different IoU thresholds, using 10 arithmetic IoU thresholds in increments of 0.05 within the range of 0.5 to 0.95. In object detection tasks, giga-floating-point operations per second (GFLOPs) are commonly used to measure the computational cost of a model.

[0135] The key points of this invention are: (1) the design of the basic network structure module for underwater target detection enhancement module inspired by biological vision mechanism; (2) the designed module is simple and efficient, and while combining attention mechanism and feature fusion, and introducing contrast enhancement, it is lightweight and plug-and-play.

[0136] This invention specifically addresses the challenges of complex underwater imaging environments, where factors such as lighting, water quality, and impurities negatively impact image quality, leading to low contrast and color variations that make feature recognition difficult. It systematically mimics the target recognition process of a higher biological visual system. First, a region perception module mimics the functions of the central and peripheral retina in a higher biological visual system. The central retina is responsible for perceiving fine target details, while the peripheral retina is used to perceive the overall scene and quickly locate the target, focusing on the target area and enhancing the extraction of key features. Second, a multi-scale feature aggregation module ensures information extraction at multiple scales, combined with an attention mechanism module to dynamically focus on important features. Finally, a contrast adaptive normalization module is added to enhance the contrast of underwater images and reduce the effects of light attenuation and blurring.

[0137] Therefore, the multi-scale feature aggregation module of this invention differs significantly from existing multi-scale fusion methods in its structural design and perception mechanism. Its fusion strategy is more biomimetic, adaptive, lightweight, and possesses global semantic understanding capabilities, making it particularly suitable for handling image scenes with complex backgrounds, uneven underwater lighting, and large target size variations. By introducing dynamic weighting and a global information perception mechanism, it effectively improves the detection capability for small and low-contrast targets—advantages not yet present or systematically integrated in existing methods. The main objective of this invention is to construct a plug-and-play, lightweight underwater target detection module driven by biological visual perception. Its design emphasizes versatility, flexibility, and deployment efficiency, serving as a functional enhancement unit for existing target detectors, improving their robustness and accuracy in complex underwater scenes without significantly increasing computational costs.

[0138] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An image enhancement device based on biomimetic vision, characterized in that: It includes a region perception module, a multi-scale feature aggregation module, a convolutional block attention mechanism module, and a contrast adaptive normalization module; The region perception module mimics the functions of the central and peripheral retina in advanced biological vision systems. The central retina is responsible for perceiving fine target details, while the peripheral retina is used to perceive the overall scene and quickly locate targets. The multi-scale feature aggregation module simulates the ability of the biological visual system to process information at different scales and depths, helping to extract and aggregate features at multiple scales, thereby comprehensively capturing image details; The convolutional block attention mechanism module simulates the attention mechanism of advanced biological visual systems, enabling it to dynamically focus on important features, suppress irrelevant information, and enhance the expressive power of key features through channel and spatial attention. The contrast adaptive normalization module simulates the mechanism of adaptive contrast adjustment in biological visual systems, automatically adjusting image brightness and contrast according to different lighting conditions to enhance the visibility of important features. The region perception module enhances the input feature map to obtain an enhanced feature map. The multi-scale feature aggregation module processes the enhanced feature map, fusing local and global features through different receptive field features to obtain fused features; the convolutional block attention mechanism module further filters the fused features to select important feature information; the contrast adaptive normalization module enhances the selected important feature information to obtain the final enhanced features; The region perception module processes the image as follows: Assuming the input feature map The size is ,in It's about height. It's the width. It refers to the number of channels; firstly, a 3×3 convolution is used to simulate the central retina's focusing on the target details, extracting local detail features, i.e. ； in, This represents a 3×3 convolution. Indicates local features, Indicates the input feature map; Next, a 5×5 convolution is used to simulate the peripheral retina's perception of the global background, extracting background features over a larger area: in, This represents a 5×5 convolution. Indicates background features, The input feature map is represented; an attention map is generated using the Sigmoid activation function. : ； in, This represents the activation function. For local features, Background features; Input feature map and attention map Element-wise multiplication is performed to enhance the features of the target region and suppress the features of the background region, resulting in an attention-weighted feature map: in, Feature maps weighted for attention, For the input feature map, For attention graphs; The processing of the multi-scale feature aggregation module is as follows: First, three depthwise separable convolutions with dilation rates of 1, 2, and 3 are used to extract feature information from different receptive fields: ； ； ； in, These represent the extracted feature information at three different scales. This represents depthwise separable convolution. Feature maps weighted for attention, Indicates the expansion rate; Next, global average pooling is performed to obtain global information from the input feature map; ； in, Represents global average pooling; This represents global information about the input feature map. Attention-weighted feature maps; Then, the pooling result is transformed into a one-dimensional feature representation, and weights at three scales are computed through a fully connected layer: in, Represents the weights of the three scales. Indicates a fully connected layer. This represents global information about the input feature map. express Activation function; Using the calculated weights Weighted fusion of multi-scale features: in, This is the weighted feature map; For expansion rate The feature weights extracted at that time, Indicates the expansion rate Feature information extracted in time; The weighted feature maps are then further fused using a 1×1 convolution, and the fused feature maps are residually connected to the input feature maps to avoid information loss. in, To fuse feature maps, This is the weighted feature map. Feature maps weighted for attention, This represents a 1×1 convolution.

2. The image enhancement device based on bionic vision according to claim 1, characterized in that: The region-aware module uses depthwise separable convolutions to replace traditional convolutions, and the 5×5 convolutions are replaced by 3×3 dilated convolutions with a dilation rate of 2.

3. The image enhancement device based on bionic vision according to claim 1, characterized in that: The convolutional block attention mechanism module includes a channel attention unit and a spatial attention unit; The main objective of the channel attention unit is to weight each channel of the feature map according to the importance of each channel, with the weight of each channel reflecting the importance of that channel in the current task. The goal of the spatial attention unit is to dynamically focus on important regions in the feature map by introducing an attention mechanism in the spatial dimension.

4. The image enhancement device based on bionic vision according to claim 3, characterized in that: The specific implementation steps of the channel attention unit are as follows: Global pooling: First, the input feature map is processed by global average pooling and global max pooling to generate a global description; global average pooling calculates the average value of each channel to capture the global information of the image; while global max pooling calculates the maximum value of each channel to capture the most prominent features. Fully connected layer: Next, the two pooling results are fed into a two-layer fully connected network; The first layer reduces the number of channels, and the second layer restores the number of channels. Each channel receives a weight representing its importance after passing through this network. Activation function: Finally, a channel-level attention weight coefficient is obtained by passing the sigmoid activation function; this weight coefficient is then multiplied channel by channel with the input feature map to obtain the adjusted feature map; in, Indicates channel attention. , These represent the weight matrices of the two fully connected layers, This represents the sigmoid activation function. Represents the ReLU activation function. Represents global average pooling. Represents global max pooling. Feature maps are fused.

5. The image enhancement device based on bionic vision according to claim 3, characterized in that: The specific implementation of the spatial attention unit is as follows: Channel operations on feature maps: By performing operations on the input feature map along the channel dimension, the importance of each position is calculated; average pooling is used to obtain global information along the channel dimension, and max pooling is used to obtain salient features along the channel dimension. Convolution operation: The feature maps of the two channels are concatenated, and a 7×7 convolution is used to extract spatial relationships over a large range, generating a spatial attention map. The output is an attention map with the same spatial dimension as the input feature map, representing the importance of each spatial location. Activation function: Finally, the values ​​of the spatial attention map are mapped to the [0,1] interval through the Sigmoid activation function, and pixel-wise multiplication is performed with the spatial dimension of the input feature map to enhance the features of important regions and suppress unimportant regions; in, Indicates channel attention, Indicating spatial attention, This represents the attention of the j-th channel. Indicates the number of channels. This represents the sigmoid activation function. This represents convolution.

6. The image enhancement device based on bionic vision according to claim 1, characterized in that: The contrast adaptive normalization module obtains the brightness information of each channel through global mean pooling, and then adaptively adjusts the image based on this brightness information. Specifically, the value of each pixel in the image is adjusted differently from the global mean of its channel. By scaling the scale and bias parameters, the adjusted image brightness adapts to different input scenarios. ； in, This indicates the final output. Indicates spatial attention. Indicates global average pooling. Indicates the scaling factor. This represents the bias parameter.

7. An underwater target detection system based on biomimetic vision, characterized in that: The system includes a target detection model comprising a backbone network, a neck network, and a head network, wherein the image enhancement device as described in any one of claims 1-6 is incorporated into the neck network.

8. An underwater target detection method based on biomimetic vision, characterized in that: The method is implemented using the underwater target detection system based on bionic vision as described in claim 7, and the method specifically includes: The raw underwater image is input into the target detection model. The backbone network of the target detection model extracts the basic features of the shallow and deep layers in the raw underwater image, and then sends them to the neck network for feature fusion. The image features are enhanced by an image enhancement device, and finally the enhanced features are input into the head network for classification and localization to perform target detection.

Images