Underwater image target detection method and device, computer equipment and storage medium

Abstract

The invention relates to an underwater image target detection method and device, computer equipment and a storage medium, and the method comprises the steps: extracting feature maps of different scales of an underwater image through a multi-scale feature extraction unit, and extracting the feature maps of different scales layer by layer through a plurality of serially-connected partial convolution re-parameterization basic blocks in the unit, in the basic block, an incomplete region in the underwater image is adaptively processed through partial convolution, and a multi-scale feature fusion unit is used for carrying out feature pyramid type cross-scale fusion operation to obtain a fusion feature map; in the unit, feature maps of different scales are subjected to global and local attention double enhancement through an enhancement block, then cross-scale feature interaction and fusion are performed through a fusion block and a convolution block, and after semantic detail repair is performed on the feature maps through a shallow repair block, the feature maps participate in the cross-scale fusion process, so that the cross-scale feature interaction and fusion are realized. And finally, detecting the underwater target by using a target detection unit according to the fused feature map. By adopting the method, the underwater target can be effectively and accurately detected.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of underwater image detection, and in particular to an underwater image target detection method and device, computer equipment and a storage medium. BACKGROUND

[0002] The ocean accounts for more than 70% of the earth's surface, and is an important natural resource, providing humans with abundant and valuable resources necessary for survival and well-being. Despite the abundance of marine resources, human understanding of the underwater world is still limited, and surveys face obstacles such as high pressure, low visibility, and drastic temperature fluctuations. Specifically, the rapid reduction of light in water can cause distortion, noise, distortion, and other problems in underwater imaging technology. Traditional marine exploration relies on divers, and all target detection and identification tasks are completed manually, which poses considerable difficulties. Advances in computer vision technology have ushered in a new era of marine exploration, promoting the deployment of non-invasive underwater exploration tools to investigate and utilize marine ecological resources. These cutting-edge technologies have been widely applied in various marine fields such as aquaculture, deep-sea fishing, marine biodiversity monitoring, environmental protection, underwater archaeology, etc. Target detection models have played an important role in this achievement.

[0003] However, compared with typical detection scenarios, underwater visual image detection presents several distinct challenges. First, the availability of underwater optical image datasets is limited, and the complex underwater ecosystem introduces noise, hindering the imaging device from capturing a large number of images. Second, underwater objects are often occluded and overlapped. In addition, underwater visual image detection encounters unique difficulties in underwater environments. These challenges can be divided into three main aspects, including: underwater images are often affected by significant noise, and often exhibit blue-green color tones, complicating the effective identification of targets. The absorption and scattering of light by the aquatic medium can cause significant interference, resulting in reduced image clarity, color distortion, and blurring. This interference enhances inter-class similarity, leading to false identification between object classes and background classes. The diversity of small underwater objects, combined with the limited availability of features, exacerbates this problem. As the network layer deepens, feature information may gradually degrade, increasing the probability of missed detection, and the multi-dimensional features of underwater objects are also evident, with inconsistent sizes, extreme aspect ratios, and diverse shapes, which can easily lead to false detection. Given these challenges, existing target detection algorithms often struggle to achieve satisfactory results in underwater target detection. SUMMARY

[0004] Therefore, it is necessary to provide an underwater image target detection method, device, computer equipment and storage medium capable of accurately detecting underwater targets in view of the above technical problems.

[0005] An underwater image target detection method, the method comprising: obtaining an underwater image; extracting feature maps of different scales of the underwater image by using a multi-scale feature extraction unit, wherein the multi-scale feature extraction unit extracts feature maps of different scales layer by layer through a plurality of serially connected partial convolution reparameterization basic blocks, and the partial convolution reparameterization basic blocks adaptively process incomplete regions in the underwater image through partial convolution; performing cross-scale fusion operation on the feature maps of different scales by using a multi-scale feature fusion unit to obtain fused feature maps, wherein, in the multi-scale feature fusion unit, the feature maps of different scales are first subjected to dual reinforcement of global and local attention through a GLSA reinforcement block, and then subjected to cross-scale feature interaction and fusion through a fusion block and a convolution block, wherein the feature map output by the last partial convolution reparameterization basic block in the multi-scale feature extraction unit is also subjected to semantic detail repair through a shallow repair block before participating in the cross-scale fusion process; detecting underwater targets according to the fused feature maps by using a target detection unit to obtain target detection results.

[0006] In one embodiment, in the partial convolution reparameterization basic block: the input data is sequentially subjected to processing of a convolution block, a batch normalization layer, and a ReLU activation function to obtain first processed data; the valid feature region of the first processed data is screened through a mask, the feature of the valid feature region is extracted by using a convolution kernel, and the extracted features are subjected to weighted summation to obtain a first summation result, the weights of the mask and the convolution kernel are associated and summed to obtain a second summation result, and the first summation result is divided by the second summation result to obtain second processed data; the first processed data and the second processed data are connected in residual, and the output data is obtained through a convolution block.

[0007] In one embodiment, in the multi-scale feature extraction unit: the preliminary shallow feature map of the underwater image is extracted through two stacked convolution blocks; the preliminary shallow feature map is subjected to step-by-step reinforcement of features through four serially connected partial convolution reparameterization basic blocks, and three feature maps of different scales are output.

[0008] In one embodiment, in the shallow repair block: candidate target positions are marked in the feature map according to preset initial sampling points; the feature map is subjected to linear transformation to generate a query vector, and the query vector is input into a learnable offset prediction network to predict a position offset; According to the predicted position offset and the candidate target position, an adjusted target position is obtained, and a relative position offset is calculated; The adjusted target position is subjected to bilinear interpolation processing to obtain a feature block aligned with the real position, and linear transformation is performed on the feature block to obtain a value vector and a key vector; The query vector, the value vector, the key vector, and the relative position offset are combined to calculate a position-aware attention weight through a multi-head attention mechanism; The feature block is aggregated using the attention weight to obtain a feature map after semantic detail repair.

[0009] In one embodiment, in the GLSA enhancement block: The input feature image is split in the channel dimension to obtain two branch features; One of the branch features is subjected to semantic enhancement processing using a global spatial attention mechanism to obtain a feature after semantic enhancement; The other branch feature is subjected to detail enhancement processing using a local spatial attention mechanism to obtain a feature after detail enhancement; The feature after semantic enhancement and the feature after detail enhancement are aggregated to obtain a deep enhanced feature.

[0010] In one embodiment, when the branch feature is subjected to semantic enhancement processing using a global spatial attention mechanism, the branch feature is sequentially subjected to convolution, transposition, softmax, residual connection, convolution, layer normalization, activation, convolution, and finally residual connection processing; When the branch feature is subjected to detail enhancement processing using a local spatial attention mechanism, the branch feature is sequentially subjected to convolution, deep convolution, residual connection, convolution, Sigmoid, element-wise multiplication, and finally residual connection processing.

[0011] In one embodiment, in the target detection unit: The target boxes in the fused feature map are filtered using IOU-aware filtering, and the target classes in each filtered target box are predicted using a multi-task detection head.

[0012] The application also provides an underwater image target detection device, which comprises: An image to be detected acquisition module is configured to acquire an underwater image; The multi-scale feature extraction module is configured to extract feature maps of different scales of the underwater image by using a multi-scale feature extraction unit, wherein the multi-scale feature extraction unit extracts feature maps of different scales layer by layer by using a plurality of serially connected partial convolution reparameterization basic blocks, and the partial convolution reparameterization basic blocks adaptively process the incomplete areas in the underwater image by using partial convolution; The cross-scale fusion module is configured to perform cross-scale fusion operation on the feature maps of different scales by using a multi-scale feature fusion unit to obtain a fused feature map, wherein in the multi-scale feature fusion unit, the feature maps of different scales are first subjected to dual reinforcement of global and local attention by using a GLSA reinforcement block, and then subjected to cross-scale feature interaction and fusion by using a fusion block and a convolution block, and wherein the feature map output by the last partial convolution reparameterization basic block in the multi-scale feature extraction unit is subjected to semantic detail repair by using a shallow repair block before participating in the cross-scale fusion process. The target detection module is configured to detect underwater targets according to the fused feature map by using a target detection unit to obtain a target detection result.

[0013] A computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the underwater image target detection method described above when executing the computer program.

[0014] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the underwater image target detection method described above.

[0015] The underwater image target detection method, device, computer device and storage medium described above extract feature maps of different scales of the underwater image by using a multi-scale feature extraction unit, wherein the multi-scale feature extraction unit extracts feature maps of different scales layer by layer by using a plurality of serially connected partial convolution reparameterization basic blocks, the partial convolution reparameterization basic blocks adaptively process the incomplete areas in the underwater image by using partial convolution, the multi-scale feature fusion unit performs cross-scale fusion operation on the feature maps of different scales to obtain a fused feature map, wherein in the multi-scale feature fusion unit, the feature maps of different scales are first subjected to dual reinforcement of global and local attention by using a GLSA reinforcement block, and then subjected to cross-scale feature interaction and fusion by using a fusion block and a convolution block, and wherein the feature map output by the last partial convolution reparameterization basic block in the multi-scale feature extraction unit is subjected to semantic detail repair by using a shallow repair block before participating in the cross-scale fusion process, and finally the target detection unit detects underwater targets according to the fused feature map. The method can effectively and accurately detect underwater targets. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 It is a flowchart of the underwater image target detection method in one embodiment; Figure 2 A structural schematic diagram of a partial convolution reparameterization base block in an embodiment; Figure 3 A structural schematic diagram of a shallow repair block in an embodiment; Figure 4 A structural schematic diagram of an FPN and a BiFPN in an embodiment; Figure 5 A structural schematic diagram of a GLSA enhancement block in an embodiment; Figure 6 A structural schematic diagram of a target detection network in an embodiment; Figure 7 Underwater image instances of targets being plastic, biomass and ROV in the Trash_ICRA19 dataset in an experiment; Figure 8 Underwater image instances of targets being sea cucumber, sea urchin, starfish and scallop in the UnderwaterBio dataset in an embodiment; Figure 9 Visualization analysis result graphs of original images, target detection using RT-DETR and the method in an experiment; Figure 10 A structural block diagram of an underwater image target detection device in an embodiment; Figure 11 An internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0018] In view of some defects of existing underwater target detection methods, an underwater image target detection method is proposed in the present application, which specifically includes the following steps: Step S100, obtaining an underwater image.

[0019] Step S110, extracting feature maps of different scales of the underwater image using a multi-scale feature extraction unit, wherein the multi-scale feature extraction unit extracts feature maps of different scales layer by layer through a plurality of serially connected partial convolution reparameterization base blocks, and the partial convolution reparameterization base block adaptively processes incomplete regions in the underwater image through partial convolution.

[0020] Step S120, a multi-scale feature fusion unit is used to perform a feature pyramid cross-scale fusion operation on the feature maps of different scales to obtain a fused feature map. In the multi-scale feature fusion unit, the feature maps of different scales are first subjected to dual strengthening of global and local attention via a GLSA strengthening block, and then subjected to cross-scale feature interaction and fusion via a fusion block and a convolution block. The feature map output by the last partial convolution reparameterization base block in the multi-scale feature extraction unit is also subjected to semantic detail repair via a shallow repair block before participating in the cross-scale fusion process.

[0021] Step S130, a target detection unit is used to detect underwater targets according to the fused feature map to obtain a target detection result.

[0022] In step S100, the underwater image is an optical image obtained underwater by a camera or a photographic equipment. Generally, underwater optical images will have degraded image quality due to optical characteristics, including light attenuation, color distortion, and scattering effects. At the same time, the background of underwater images is complex, and the target scales differ greatly. The underwater scene background is rich and varied, including reefs, seaweed, bubbles, plankton, etc. These background elements and target objects are intertwined with each other, which easily forms interference, causing the neural network to mistakenly regard the background as a target or miss the real target. For example, in an area with dense seaweed, fish may be detected, and the shape and color of the seaweed may be similar to those of certain fish, resulting in detection errors. The sizes of underwater targets differ significantly, from tiny plankton to large whales, sunken ships, etc. This puts high requirements on the scale adaptability of the neural network, and a single network structure often cannot simultaneously detect targets of different scales, easily leading to problems such as missing small targets and low detection accuracy for large targets.

[0023] In the present application, a neural network structure for underwater target detection is proposed, which includes the multi-scale feature extraction unit in step S110, the multi-scale feature fusion unit in step S120, and the target detection unit in step S130.

[0024] In step S110, the multi-scale feature extraction unit is improved based on the traditional ResNet architecture. Although the BasicBlock structure in the original backbone has strong expression ability, it has certain limitations in specific tasks, especially in the underwater biological target detection scene. Underwater images often have problems such as reduced contrast, blurring, and noise, which hinder the effective extraction of basic features by conventional convolution operations. In addition, the BasicBlock in the traditional ResNet uses a fixed convolution structure, which is difficult to adapt to the complex changes and low-quality feature inputs of underwater images. Therefore, in the present embodiment, a partial convolution reparameterization base block, BasicBlock_PConv_Rep, is proposed, as shown in Figure 2As shown, the basic block combines partial convolution (PConv) and reparameterization (Rep) techniques, thereby enhancing the model's ability to extract incomplete regions (such as noise or missing parts) and complex features in underwater images. BasicBlock_PConv_Rep not only improves feature representation, but also significantly reduces computational requirements and memory consumption. With this enhancement, the model can better detect underwater organisms, overcoming the limitations of traditional BasicBlock in such tasks.

[0025] In this embodiment, the partial convolution reparameterization basic block extends the traditional BasicBlock by incorporating a new module PartialConv3_Rep as its basic component. The main purpose of this module is to adaptively process incomplete regions of the image through PConv, thereby improving the model's performance on underwater images. PConv effectively addresses missing parts of the input image by utilizing a mask to process only the valid regions for convolution operations. This approach alleviates the interference caused by traditional convolution methods in the presence of noise.

[0026] Specifically, in the partial convolution reparameterization basic block, the input data is first processed by a convolution block, a batch normalization layer, and a ReLU activation function in sequence to obtain first processed data. Then, the valid feature regions of the first processed data are screened using a mask, and the features of the valid feature regions are extracted using a convolution kernel. The extracted features are then weighted and summed to obtain a first summation result. The mask and the weights of the convolution kernel are associated and summed to obtain a second summation result. The first summation result is divided by the second summation result to obtain second processed data. Finally, the first processed data and the second processed data are connected in residual, and the output data is obtained through a convolution block.

[0027] Further, the operation of partial convolution is mathematically represented as:

[0028] In the above formula, represents the first processed data, represents the mask, represents the convolution kernel, represents the dot product. When the mask indicates a specific condition, the convolution operation ignores the corresponding region. Only when the condition is met, it indicates that the corresponding pixel in the feature map is valid and should be included in the convolution calculation. This approach significantly reduces the impact of noise or missing parts in the image on the resulting features. In addition, the Rep operation is also adopted to improve the performance of the convolution layer, making the convolution process more effectively adapt to the challenging task of detecting underwater targets.

[0029] In this embodiment, based on the above-mentioned partial convolution reparameterization basic block, a multi-scale feature extraction unit is further proposed. In the unit, two stacked convolution blocks are used to extract preliminary shallow feature maps of the underwater image, and the convolution kernel size can be set to 3X3. Then, the preliminary shallow feature maps are gradually strengthened through four partial convolution reparameterization basic blocks connected in series, and three feature maps of different scales are output, which are the first feature map, the second feature map and the third feature map from shallow to deep.

[0030] In step S120, in the multi-scale feature fusion unit, a shallow repair block, namely Da-AIFI block, is designed. In the traditional architecture, the adaptive information fusion and interaction (AIFI) layer is usually used to manage the feature fusion and interaction of each level. However, the AIFI block has certain limitations in solving long-range dependencies, global information modeling and multi-scale feature fusion, especially in complex scenes such as underwater biological detection. It cannot fully utilize the global context information to limit the performance of the model. In order to overcome this challenge, in this embodiment, a Da-AIFI block is proposed, and its structure is shown in Figure 3 The block is based on the architecture of the Transformer encoder and combines the deformable attention mechanism, which makes great progress in the representation of global context information. By dynamically adjusting the key area of attention, the Da-AIFI block not only improves the accuracy of object detection in underwater biological monitoring tasks, but also skillfully handles long-range dependencies and complex backgrounds, significantly improving the flexibility of the model.

[0031] Specifically, Da-AIFI adopts an enhanced Transformer encoder architecture, and the most significant improvement is the introduction of a deformable attention mechanism. This mechanism adaptively and differently focuses on the local and global aspects of image features. In Da-AIFI, an attention mechanism process combining position information and feature information is proposed: first, a reference point generates a query vector (q), and the geometric offset is calculated to form a "position flow" that describes the spatial relationship; at the same time, the image features in the "feature flow" are combined with the relative position offset and bilinear interpolation to generate position-aware feature vectors v~ and k~; finally, the position information dynamically calculated by the position flow and the image information obtained by the feature flow are input into the multi-head attention module for fusion processing, so that the model can utilize both position geometric information and input feature information, and finally generate an output representation Z with spatial perception ability.

[0032] Specifically, the Da-AIFI (shallow repair block) is arranged in the multi-scale fusion unit, and the third feature map is repaired. In the shallow repair block: according to a preset initial sampling point in the feature map, a candidate target position is marked, a linear transformation is performed on the feature map to generate a query vector, the query vector is input into a learnable offset prediction network, a position offset is predicted, an adjusted target position is obtained according to the predicted position offset and the candidate target position, and a relative position offset is calculated. The adjusted target position is bilinearly interpolated to obtain a feature block aligned with the real position. The feature block is linearly transformed to obtain a value vector and a key vector. The query vector, the value vector, the key vector and the relative position offset are combined to calculate the position-aware attention weight through the multi-head attention mechanism. Finally, the feature block is aggregated by using the attention weight to obtain a feature map after semantic detail repair.

[0033] In the embodiment, it is considered that in the traditional cross-scale feature fusion unit, a single convolution operation and a feature fusion strategy are mainly used, for example, a widely used feature pyramid network (FPN). However, this structure shows limitations in managing complex tasks, especially in scenes such as underwater object detection, in which multi-scale changes, complex backgrounds and low contrast conditions are common. These factors often hinder the effective integration of multi-scale information and global context. For underwater detection tasks, the traditional FPN and single convolution technology often have difficulty in fully addressing the inherent noise and complex background in underwater images, resulting in reduced detection accuracy. In order to overcome these challenges, in the embodiment, a G-Bifpn multi-scale fusion structure is proposed, which aims to alleviate the limitations of traditional methods in underwater detection, as shown in Figure 4 As shown in Figure 4 , the BiFPN effectively combines features from various scales using bidirectional information flow, covering top-down and bottom-up paths. It performs weighted fusion at each layer, ensuring efficient information transmission and lossless integration. Its workflow is represented as:

[0034] In the above formula, represents the feature map of the i-th scale, is the weight associated with each scale feature, and N represents the number of feature scales, is a parameter automatically adjusted during training.

[0035] In the embodiment, when the feature maps of different scales are interactively fused in the multi-scale feature fusion unit, the GLSA enhancement block is first used to enhance each feature map to achieve two main goals: improving the accuracy of the previous features during feature fusion and suppressing the propagation of irrelevant formations, and minimizing noise artifacts. ​like Figure 5 As shown in the figure, in the GLSA enhancement block: the input feature image is split into two branch features in the channel dimension. One of the branch features is semantically enhanced using the global spatial attention mechanism (GSA mechanism) to obtain the semantically enhanced feature. The other branch feature is processed using the local spatial attention mechanism (LSA) to enhance the details to obtain the detail enhanced feature. The semantically enhanced feature and the detail enhanced feature are aggregated to obtain the deep enhanced feature. The data processing process can be expressed as:

[0036] In the above formula, and In order to obtain two branch features after splitting the feature image, are semantically enhanced features and detail enhanced features respectively.

[0037] Specifically, the Global Spatial Attention (GSA) mechanism enhances the grasp of global context by establishing long-range associations, thereby enhancing global semantic content in low-level features. The Local Spatial Attention (LSA) mechanism improves the accuracy of detecting small underwater objects by focusing on specific regions of interest, thereby preserving and enhancing complex details of low-level features.

[0038] Furthermore, when the global spatial attention mechanism is used to perform semantic enhancement on the branch features, the branch features are sequentially processed with convolution, transposition, softmax, residual connection, convolution, layer normalization, activation, convolution, and finally residual connection. When the local spatial attention mechanism is used to perform detail enhancement on the branch features, the branch features are sequentially processed with convolution, depthwise convolution, residual connection, convolution, sigmoid, element-wise multiplication, and finally residual connection. The structure is referenced Figure 5 .

[0039] In this embodiment, the multi-scale fusion unit uses a GLSA enhancement block to extract more expressive local and global features from feature maps at different scales. For each scale feature map, the GLSA enhancement block processes local and global information separately to generate deep enhanced features that retain detailed information and contextual information about the target. Subsequently, the deep enhanced features are input into the BiFPN structure, which performs a weighted integration of features at various scales to generate the final multi-scale fused features. The integrated G-BiFPN module greatly improves the accuracy and robustness of underwater target detection, especially in complex environments with blurred targets and significant scale variations.

[0040] In step S130, in the target detection unit: first, the target boxes in the fused feature map are screened using IOU perception screening, and then the target classes in each screened target box are predicted using the multi-task detection head.

[0041] In the present embodiment, a new target detection network is constructed according to the above-mentioned multi-scale feature extraction unit, multi-scale feature fusion unit and target detection unit. In the target detection network, the synergistic effect among the newly proposed partial convolution reparameterization basic block (BasicBlock_PConv_Rep), GLSA enhancement block and shallow repair block enables the target detection network to accurately detect underwater images and effectively cope with underwater scenes.

[0042] Specifically, the partial convolution reparameterization basic block, as an enhanced convolution module, replaces the traditional BasicBlock. When the local region of the image is missing or incomplete, the partial convolution reparameterization basic block uses the partial convolution technology to restore the information by filling the existing region features, and at the same time, further perfects the structure description of the feature map by merging the Rep block, so that the network can process local detailed information with higher efficiency. These enhancements enable the network to more robustly handle regions with incomplete or missing information, thereby improving the effectiveness and robustness of the model.

[0043] Specifically, the shallow repair block (Da-AIFI) integrates a deformable attention mechanism that selectively focuses on regions of interest. This enables the target detection network to maintain a high level of attention on key image features while reducing computational load. Therefore, this enhancement not only greatly improves the model's ability to capture global information, but also enhances its robustness, enabling it to more effectively cope with various challenges in underwater biological detection.

[0044] Specifically, in the multi-scale feature fusion unit, after combining the GLSA enhancement block with the Bifpn architecture, the model is allowed to efficiently handle different scales of functions while minimizing computational requirements, thereby improving accuracy in complex scenarios. The G-Bifpn architecture improves the performance of object detection models in underwater images by integrating GLSA and Bifpn architectures, which facilitates the processing of local and global features. In addition, the bidirectional information flow and weighted aggregation mechanism further enhance the model's detection capabilities, significantly improving the accuracy and robustness of underwater object detection, especially in complex environments where target blur and scale variation are significant.

[0045] In one embodiment, the structure of the entire target detection network is as shown in Figure 6 .

[0046] In this paper, the effectiveness of the method is also demonstrated through experiments.

[0047] The datasets used in the experiments are Trash_ICRA19 and UnderwaterBio datasets. Among them, the Trash_ICRA19 dataset is an important open source tool for identifying underwater objects in marine environments, carefully annotated in PASCAL VOC format. This collection features images of different quality, depth, and object conditions, showing marine litter, various organisms, and other artifacts. The dataset contains a total of 5720 images. Comprehensive analysis shows that the subset consists of 853 plastic fragments, 70 biological samples, and 141 ROV units, as shown in Figure 7 In addition, the test set includes 937 pieces of plastic, 396 biological entities, and 335 ROV units, providing rich data for further testing. UnderwaterBio data is a multi-class image dataset for underwater object detection and classification, containing 5454 images, including 544 test images and 1095 validation images, each with a resolution of 640 x 640 pixels. The dataset aims to identify marine biological objects, covering four main categories: holothurians, echinus, scallop, and starfish. The dataset provides a variety of challenges as an important experimental basis for underwater image processing research, especially for image classification, object detection, and environmental context understanding activities. It effectively tests and evaluates the performance of various visual models. Figure 8 A subset of the dataset is shown.

[0048] Then, Tables 1 and 2 respectively give the hardware platform and environmental parameters used in the experimental training stage, and the key parameters set during the training of the model (target detection network).

[0049] Table 1 Hardware platform and environmental parameters used in the experimental training stage

[0050] Table 2 Key parameters set during the training of the model (target detection network)

[0051] During the experiment, mAP, recall (R) and precision (P) were used as quantitative indicators to evaluate the effectiveness of the proposed method. mAP is used as a standard for evaluating target detection algorithms, calculated by DOTA metrics. Precision evaluates the model's ability to accurately identify relevant objects, representing the proportion of correct predictions to the model's total predictions. At the same time, Recall measures the model's ability to detect all relevant objects, representing the maximum number of actual objects that the model's predictions can cover. The formulas for Precision, Recall, and Average Precision (AP) are:

[0052]

[0053] In the above formula, TP (True Positive) represents the number of samples that are accurately predicted as positive, FP (False Positive) represents the count of TP / (TP negative + FN) samples that are recalled as positive (13) that are wrongly predicted, and FN represents the number of positive samples that are ignored.

[0054] mAP@0.5 refers to the mean average precision evaluated at a cross point / join (inter-distance) threshold of 0.50. In contrast, mAP@0.5-0.95 refers to the mean average precision evaluated over a range of inter-distance thresholds, from 0.50 to 0.95, in increments of 0.05. The inter-distance measures the ratio of the overlapping area between a true bounding box and a predicted bounding box to the total area covered by both bounding boxes. To find the inter-distance value, one must first determine the inter-distance of each predicted bounding box with respect to the ground truth box. Following this, the computed inter-distance values can be used to construct a precision-recall (PR) curve. Precision is defined as the ratio of true positives to the total number of positive predictions, while Recall is described as the ratio of true positives to the actual number of positive examples. The average precision (AP) is obtained by computing the area under the PR curve for each individual class. The AP reflects the average precision of the model’s predictions for a particular class. The overall mAP (mean average precision) is established by averaging the AP values for all classes, thereby providing an overall evaluation of the model’s target detection capabilities across different classes. The formula for AP and mAP is represented as:

[0055]

[0056] In the above formula, P(i) represents the recall precision corresponding to the ith recall level, dR(i) reflects the change in recall rate between consecutive levels, represents the average recall precision for the jth class, and N represents the total count of queries or classes.

[0057] The images in the Trash_ICRA19 dataset contain a variety of complex marine targets and environmental changes in different backgrounds, showcasing the diversity of underwater marine trash and biological species. This provides a rich and diverse data set for model training. In the underwater environment, the high similarity between targets and backgrounds increases the difficulty of detection algorithms, making it difficult for the model to accurately distinguish targets. However, this challenge also puts a strict test on the detection ability of the model, promoting the improvement of the model's robustness and accuracy in complex environments. During training, the loss and accuracy change is quite obvious. In the first few iterations, the loss decreases significantly, while the accuracy improves significantly. As the iterations proceed, the loss gradually decreases and the precision slightly improves, and the model eventually converges. The overall Precision and Recall are 0.99 and 0.971, respectively. As shown in the figure, the performance of the three classes is very good, among which Plastic achieves the best detection performance with a recall rate of 0.985. This is due to the significant color contrast between Plastic and the background, as well as the ease of detecting medium-sized objects. In addition, it is clear that the size of the dataset has an impact on the performance of the model. From the mAP value, it is observed that the plastic, ROV and bio categories all exceed 0.95, and the overall class precision reaches 0.987. The mAP@0.5:0.95 precision reaches 0.821. As shown in Table 3, the target detection in this method performs well in all categories. Although the model is still slightly affected by the background, it shows strong classification ability, especially in complex backgrounds. This further verifies the robustness and efficiency of the model in practical applications. The existing models have not yet reached such accuracy, which highlights the superiority of this method.

[0058] Table 3 Performance of target detection network in this method in different categories

[0059] The UnderwaterBio dataset provides rich and challenging data for model training, especially in the underwater environment where target objects are very similar to the background and usually very small. Throughout the training process, the model's performance gradually improves. The training data shows that the model gradually performs better in all target detection tasks, especially in terms of Precision and Recall. The detection results of each class are also impressive, highlighting the model's strong adaptability and stability in handling multi-class detection tasks. Finally, on the test set, the model achieved an mAP@0.5 score of 0.792 and an mAP@0.5:0.95 score of 0.434. The results shown in Table 4 demonstrate the model's high efficiency and reliability in complex multi-class detection tasks, effectively solving the challenging target detection problem in underwater environments.

[0060] Table 4 Performance of the target detection network in the present method in different categories

[0061] Further, in order to prove the superiority of the improved model, a series of extensive benchmark evaluations were also conducted, and its performance was compared with several contemporary target tracking models. These models include the resource-intensive Faster R-CNN, the highly skilled SSD, and the compact but powerful YOLOv7-tiny. In addition, the model of the present method was evaluated with various adaptations of YOLOv8, such as YOLOv8-goldyolo, YOLOv8-convnextv2, YOLOv8-swintransformer, and YOLOv8-vanillanet, as well as the improved YOLOv8 models developed by DETR, RT-DETR, and Chhaya Gupta. Table 5 summarizes the comparison results on the Trash_ICRA19 dataset. The detection framework proposed by the present method achieved an mAP@0.5 of 0.987, which significantly exceeded the performance of all other models, highlighting the considerable performance advantage brought by the design of the present model.

[0062] In addition, comparative experiments were also conducted with several well-known target detection models, and the research results are shown in Table 5. Then, compared with Faster R-CNN (mAP@0.5 = 0.712), SSD, YOLOv7 series (represented by YOLOv7, mAP@0.5 = 0.823), and DETR, the model proposed by the present method showed superior performance in various indicators. Notably, the model of the present method achieved an mAP@0.5 = 0.987, significantly better than all other comparison models. In addition, compared with YOLOv8-goldyolo (mAP@0.5 = 0.832), the model of the present method significantly improved the target detection accuracy while maintaining a relatively low computational complexity. Furthermore, the model proposed by the present method performed well in precision and recall, reaching 0.99 and 0.971, respectively, better than most of the compared models. Finally, compared with the initial model, the computational load of the model of the present method was significantly reduced.

[0063] Table 5 Experimental results on the Trash_ICRA19 dataset

[0064] Further, the performance of the target detection model proposed in this method on the underwater biological data set is evaluated, and it is compared with existing mainstream target detection models. Faster R-CNN, SSD, YOLOv5-DETR, RT-DETR and YOLOv8-DETR are selected as benchmarks, and precision, recall, mAP@0.5, mAP@0.5-0.95 and computational cost (GFLOPs) are used as evaluation indicators.

[0065] Table 6 shows the comprehensive evaluation results of each model. As shown in Table 6, although Faster R-CNN has a high recall rate (0.691), its precision (0.444) and mAP values (mAP@0.5 = 0.52, mAP@0.5-0.95 = 0.236) are relatively low, and the computational cost is huge (940.9 GFLOPs). This shows that it has certain limitations when dealing with complex underwater biological images. The SSD model has high precision (0.677), low recall (0.447), and low mAP@0.5 (0.518). Although the computational cost is moderate (62.7 GFLOPs), its detection in complex background needs to be improved. The YOLOv5-DETR model achieves a harmonious balance between precision (0.762) and recall (0.663), with an mAP@0.5 score of 0.721 and a reasonable computational cost of 10.7 GFLOPs. Although the YOLOv8-DETR model has marginal improvements in precision (0.774) and mAP@0.5 (0.718), the recall rate (0.657) is comparable, and it also maintains a reasonable cost, but the model proposed in this method is always superior to both in all evaluation indicators. The RT-DETR model has significantly improved accuracy (0.789) and recall (0.719), with a recall rate mAP@0.5 of 0.773, showing strong detection ability. However, the model proposed in this method has a more reasonable computational cost (36.4 GFLOPs) while maintaining high accuracy and recall, indicating its advantages in efficiency and practicality. It can be seen that the target detection model proposed in this method achieves the best performance in terms of precision (0.792), recall (0.725), mAP@0.5 (0.78) and mAP@0.5-0.95 (0.434). At the same time, its computational cost (36.4 GFLOPs) is within a reasonable range, indicating that the model has high precision and efficiency when dealing with complex underwater biological images, and has good potential for practical application.

[0066] Table 6 Experimental results on UnderwaterBio dataset

[0067] Further, comprehensive ablation experiments were conducted to investigate the individual contribution of each unit to the performance of the target detection model. The experimental results are shown in Table 7, including precision, recall, GFLOPs (a measure of computational complexity), parameter count (Param / MB), and mAP@0.5. The baseline model has a precision of 0.921, a recall of 0.881, an mAP@0.5 of 0.738, a computational cost of 58.6 GFLOPs, and a parameter of 20.8 MB. After introducing the DA-AIFI module, although the computational cost increased slightly, both the precision (to 0.978) and the recall (to 0.956) were significantly improved. In contrast, the integration of the G-BiFPN module not only slightly improved the precision (to 0.974), but also reduced the computational cost and parameters, highlighting its efficiency in improving model performance without increasing computational burden. In addition, the introduction of the P-Rep module brought significant improvement, with the precision further improved to 0.987 and the recall to 0.943, while reducing the computational complexity (to 44.6 GFLOPs) and the parameters (to 15.0 MB). The final combined model, including the baseline model, DA-AIFI, G-BiFPN, and P-Rep, achieved an amazing precision of 0.99, a recall of 0.971, an mAP@0.5 of 0.987, a reduction in computational cost of 36.4 GFLOPs, and a reduction in parameters of 12.5 MB, demonstrating the best balance between performance and efficiency. These results highlight the substantial contribution of each module and the enhanced detection ability and efficiency of the integrated model compared to the baseline.

[0068] Table 7 Ablation experiment results

[0069] After analyzing the accuracy of each unit, visual detection boxes and classification were used on the Trash_ICRA19 dataset for analysis and comparison, as shown in Table 8. Figure 9 As shown in Table 8, the precision of the baseline model is 0.921, the recall is 0.881, and the mAP@0.5 is 0.738. After introducing the DA-AIFI module, the precision is improved to 0.978 and the recall to 0.956. In contrast, the integration of the G-BiFPN module not only slightly improves the precision (to 0.974), but also reduces the computational cost and parameters, highlighting its efficiency in improving model performance without increasing computational burden. In addition, the introduction of the P-Rep module brings significant improvement, with the precision further improved to 0.987 and the recall to 0.943, while reducing the computational complexity (to 44.6 GFLOPs) and the parameters (to 15.0 MB). The final combined model, including the baseline model, DA-AIFI, G-BiFPN, and P-Rep, achieves an amazing precision of 0.99, a recall of 0.971, an mAP@0.5 of 0.987, a reduction in computational cost of 36.4 GFLOPs, and a reduction in parameters of 12.5 MB, demonstrating the best balance between performance and efficiency. These results highlight the substantial contribution of each module and the enhanced detection ability and efficiency of the integrated model compared to the baseline. Figure 9As shown, the model not only accurately identifies the category but also has high recognition accuracy for underwater objects. Taking ROV targets as an example, the baseline RT-DETR model can perform category recognition; but there are still false positives, missed small targets, and repeated detection boxes in the identification process. The model proposed in this application significantly outperforms the baseline model in detection performance, skillfully identifying smaller targets, thus demonstrating its superior feature extraction capability. Similarly, in plastic detection, the model proposed in this application successfully detects the targets missed by the baseline RT-DETR model, indicating that its ability to detect smaller targets has been significantly enhanced, and the accuracy has also improved. Finally, in mixed biological and plastic images, the baseline model incorrectly identifies biological as plastic, and some biological detection boxes exhibit redundancy and repetition, with low confidence scores for the identified category. The model proposed in this method solves these problems and clearly displays the detected targets, highlighting its advantages in feature extraction and processing capabilities.

[0070] In the above-mentioned underwater image target detection method, a deep learning-based underwater target detection method is proposed, which solves the inherent complexity and challenges of detecting objects in underwater environments. To solve this problem, a novel feature extraction module is proposed, which combines partial convolution with repeated convolution (partial convolution reparameterization base block), thereby enhancing the feature extraction capability of the model while minimizing resource consumption. In addition, an improved version of the Transformer encoder architecture is proposed, which ingeniously integrates a deformable attention mechanism that allows simultaneous attention to local and global information in the image. This enhanced design significantly improves the processing capability of the model when processing images with complex spatial dependencies and multi-scale features, thereby improving its performance in underwater target detection tasks (GLSA enhancement block). In addition, a spatial pyramid structure is also introduced to further enhance the feature extraction capability of the cross-scale feature fusion unit. The effectiveness of the model is verified through ablation experiments on the public dataset Trash_ICRA19. The proposed model is compared with various state-of-the-art target detection models on the Trash_ICRA19 and UnderwaterBio datasets, confirming its effectiveness. Experimental results show that the model proposed in this method has achieved substantial performance enhancement in mAP@0.5 and mAP@0.5-0.95 indicators.

[0071] It should be understood that, although Figure 1 The steps in the flowchart of the method of the present application are shown in sequence according to the direction of the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated otherwise herein, there is no strict order limitation for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1At least one of the steps in the method can comprise a plurality of sub-steps or a plurality of stages, which are not necessarily performed at the same time, but can be performed at different times, and the order of the execution of the sub-steps or stages is not necessarily sequential, but can be performed alternately or alternately with at least one of the other steps or sub-steps or stages of the other steps.

[0072] In one embodiment, as shown in Figure 10 An underwater image target detection device is provided, comprising: an image to be detected acquisition module 200, a multi-scale feature extraction module 210, a cross-scale fusion module 220 and a target detection module 230, wherein: The image to be detected acquisition module 200 is configured to acquire an underwater image. The multi-scale feature extraction module 210 is configured to extract feature maps of different scales of the underwater image by using a multi-scale feature extraction unit, wherein the multi-scale feature extraction unit extracts feature maps of different scales layer by layer by a plurality of serially connected partial convolution reparameterization basic blocks, and the partial convolution reparameterization basic blocks adaptively process incomplete regions in the underwater image by partial convolution. The cross-scale fusion module 220 is configured to perform a cross-scale fusion operation on the feature maps of different scales by using a multi-scale feature fusion unit to obtain a fused feature map, wherein in the multi-scale feature fusion unit, the feature maps of different scales are first subjected to dual reinforcement of global and local attention by a GLSA reinforcement block, and then subjected to cross-scale feature interaction and fusion by a fusion block and a convolution block, and wherein the feature map output by the last partial convolution reparameterization basic block in the multi-scale feature extraction unit is also subjected to semantic detail repair by a shallow repair block before participating in the cross-scale fusion process. The target detection module 230 is configured to detect underwater targets according to the fused feature map by using a target detection unit to obtain a target detection result.

[0073] For specific limitations of the underwater image target detection device, please refer to the limitations of the underwater image target detection method described above, which will not be repeated here. Each module in the above underwater image target detection device can be realized by software, hardware and their combinations. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0074] In one embodiment, a computer device is provided, which can be a terminal, and its internal structure diagram can be as shown in Figure 11As shown in the figure. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected by a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface of the computer device is used to communicate with the external terminal through the network connection. The computer program is executed by the processor to implement an underwater image target detection method. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0075] Those skilled in the art can understand that, Figure 11 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0076] In one embodiment, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the following steps: Obtaining an underwater image; Extracting feature maps of different scales of the underwater image by using a multi-scale feature extraction unit, wherein the multi-scale feature extraction unit extracts feature maps of different scales layer by layer through a plurality of serially connected partial convolution reparameterization basic blocks, and the partial convolution reparameterization basic blocks adaptively process incomplete regions in the underwater image through partial convolution; Performing a feature pyramid cross-scale fusion operation on each feature map of different scales by using a multi-scale feature fusion unit to obtain a fused feature map, wherein in the multi-scale feature fusion unit, the feature maps of different scales are first subjected to dual reinforcement of global and local attention by a GLSA reinforcement block, and then subjected to cross-scale feature interaction and fusion by a fusion block and a convolution block, wherein the feature map output by the last partial convolution reparameterization basic block in the multi-scale feature extraction unit is also subjected to semantic detail repair by a shallow repair block before participating in the cross-scale fusion process; Detecting underwater targets according to the fused feature map by using a target detection unit to obtain a target detection result.

[0077] In one embodiment, a computer readable storage medium is provided, having stored thereon a computer program which, when executed by a processor, implements the following steps: obtaining an underwater image; extracting feature maps of different scales of the underwater image by using a multi-scale feature extraction unit, wherein the multi-scale feature extraction unit extracts feature maps of different scales layer by layer through a plurality of serially connected partial convolution reparameterization base blocks, and the partial convolution reparameterization base blocks adaptively process incomplete regions in the underwater image through partial convolution; performing cross-scale fusion operation on the feature maps of different scales by using a multi-scale feature fusion unit to obtain a fused feature map, wherein in the multi-scale feature fusion unit, the feature maps of different scales are first subjected to dual reinforcement of global and local attention by a GLSA reinforcement block, and then subjected to cross-scale feature interaction and fusion by a fusion block and a convolution block, and wherein the feature map output by the last partial convolution reparameterization base block in the multi-scale feature extraction unit is also subjected to semantic detail repair by a shallow repair block before participating in the cross-scale fusion process; detecting underwater targets according to the fused feature map by using a target detection unit to obtain a target detection result.

[0078] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM) and memory bus dynamic RAM (RDRAM).

[0079] Any combination of the technical features in the above embodiments can be made, and for the sake of brevity, not all possible combinations are described above, however, as long as the combination of the technical features does not exist in contradiction, it shall be considered within the scope of the present disclosure.

[0080] The above embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it shall not be understood as a limitation on the patent scope of the present application. It shall be pointed out that, for ordinary skilled persons in the art, several modifications and improvements can be made without departing from the concept of the present application, and these shall be within the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A method for underwater image target detection, characterized in that: The method comprises: Acquire underwater images; Extracting feature maps of different scales of the underwater image using a multi-scale feature extraction unit, wherein the multi-scale feature extraction unit extracts feature maps of different scales layer by layer through a plurality of serially connected partial convolution re-parameterized basic blocks, and the partial convolution re-parameterized basic blocks adaptively processes incomplete areas of the underwater image through partial convolution; A multi-scale feature fusion unit is used to perform a feature pyramid-style cross-scale fusion operation on feature maps of different scales to obtain a fused feature map. In the multi-scale feature fusion unit, the feature maps of different scales are first subjected to dual enhancement of global and local attention via a GLSA enhancement block, and then to cross-scale feature interaction and fusion via a repass fusion block and a convolution block. The feature map output by the last partial convolution re-parameterized basic block in the multi-scale feature extraction unit is further subjected to semantic detail restoration via a shallow restoration block before participating in the cross-scale fusion process. The target detection unit is used to detect underwater targets according to the fusion feature map to obtain a target detection result.

2. The underwater image target detection method according to claim 1, characterized in that: In the partial convolution reparameterization basic block: The input data is first processed by the convolution block, batch normalization layer and ReLU activation function in sequence to obtain the first processed data; screening a valid feature region of the first processed data using a mask, extracting features of the valid feature region using a convolution kernel, performing weighted summation on the extracted features to obtain a first summation result, associating the mask and the weights of the convolution kernel and then summing them to obtain a second summation result, and dividing the first summation result by the second summation result to obtain second processed data; After performing residual connection on the first processed data and the second processed data, output data is obtained through a convolution block.

3. The underwater image target detection method according to claim 2, characterized in that: In the multi-scale feature extraction unit: extracting a preliminary shallow feature map of the underwater image through two stacked convolution blocks; The features of the preliminary shallow feature map are then gradually enhanced through four partially convolutional re-parameterized basic blocks connected in series, and three feature maps of different scales are output.

4. The underwater image target detection method according to claim 2, characterized in that: In the shallow repair block: Marking candidate target locations in the feature map according to a preset initial sampling point; Performing a linear transformation on the feature map to generate a query vector, inputting the query vector into a learnable offset prediction network to predict a position offset; According to the predicted position offset and the candidate target position, the adjusted target position is obtained and the relative position offset is calculated; Performing bilinear interpolation processing on the adjusted target position to obtain a feature block aligned with the real position, and performing linear transformation on the feature block to obtain a value vector and a key vector; Combining the query vector, the value vector, the key vector, and the relative position offset, a position-aware attention weight is calculated using a multi-head attention mechanism; The feature blocks are aggregated using the attention weights to obtain a feature map after semantic detail restoration.

5. The underwater image target detection method according to claim 2, characterized in that: In the GLSA hardening block: Split the input feature image into channel dimensions to obtain two branch features; A global spatial attention mechanism is used to perform semantic enhancement on one of the branch features to obtain the semantically enhanced feature; The local spatial attention mechanism is used to enhance the details of the other branch feature to obtain the feature after detail enhancement; The semantically enhanced features and the detail enhanced features are aggregated to obtain deep enhanced features.

6. The underwater image target detection method according to claim 5, characterized in that: When the global spatial attention mechanism is used to perform semantic enhancement processing on the branch features, the branch features are sequentially subjected to convolution, transposition, softmax, residual connection, convolution, layer normalization, activation, convolution, and finally residual connection processing; When the branch features are processed for detail enhancement using the local spatial attention mechanism, the branch features are sequentially processed with convolution, depthwise convolution, residual connection, convolution, sigmoid, element-wise multiplication, and finally residual connection.

7. The underwater image target detection method according to claim 2, characterized in that: In the target detection unit: The target boxes in the fused feature map are screened using IOU-aware filtering, and then the target categories in each screened target box are predicted using a multi-task detection head.

8. An underwater image target detection device, characterized in that: The device comprises: An image acquisition module to be detected, used to acquire underwater images; a multi-scale feature extraction module, configured to extract feature maps of different scales of the underwater image using a multi-scale feature extraction unit, wherein the multi-scale feature extraction unit extracts feature maps of different scales layer by layer through a plurality of serially connected partial convolution re-parameterized basic blocks, and the partial convolution re-parameterized basic blocks adaptively process incomplete areas of the underwater image through partial convolution; A cross-scale fusion module is used to perform a feature pyramid-style cross-scale fusion operation on feature maps of different scales using a multi-scale feature fusion unit to obtain a fused feature map. In the multi-scale feature fusion unit, feature maps of different scales are first subjected to dual enhancement of global and local attention via a GLSA enhancement block, followed by cross-scale feature interaction and fusion via a repass fusion block and a convolution block. The feature map output by the last partial convolution re-parameterized basic block in the multi-scale feature extraction unit is further subjected to semantic detail restoration via a shallow restoration block before participating in the cross-scale fusion process. The target detection module is used to detect underwater targets using the target detection unit according to the fusion feature map to obtain target detection results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images