Improved method for high-precision target detection in marine fisheries based on RT-DETR

Abstract

The present invention discloses an improved method for high-precision target detection in marine fisheries based on RT-DETR, comprising collecting and preprocessing multi-perspective images of marine organisms to obtain a marine fishery target dataset, using an Efficient Block module to replace the Basic Block module in the backbone network of the model, introducing a cascaded grouped convolutional attention module to replace the AIFI module in the model, introducing a cross-scale fusion module based on an attention mechanism into the neck network of the model, configuring a training environment to train the improved method using the input of the marine fishery target dataset, and using the target image to be detected as input to obtain marine fishery target detection results and marine fishery target recognition information as output. The present invention achieves a good balance between target recognition accuracy and network volume, improves positioning accuracy and recognition accuracy, and can enhance detection capabilities in complex marine fishery scenarios.

Description

Technical Field

[0001] The present invention relates to an improved method for high-precision target detection in marine fisheries, specifically to an improved method for high-precision target detection in marine fisheries based on RT-DETR, and belongs to the technical field of target detection algorithms based on deep learning. Background Art

[0002] Generally speaking, the goal of achieving a smart ocean and smart aquaculture is to create an information-based ocean by integrating the ocean with artificial intelligence and building an ocean information system centered on intelligent marine information infrastructure. The concept of a smart ocean is driving the development of information-based aquaculture in marine fisheries, marine ranching, and mariculture, promoting the transformation and upgrading of marine agriculture and animal husbandry.

[0003] Currently, many marine ranching companies are using underwater camera systems to collect and utilize video data. The sheer volume of real-time video and the amount of information available make it difficult to extract biome data from this massive amount of video data manually. Therefore, artificial intelligence and image analysis technologies are becoming increasingly important in marine ranching bioresource monitoring research. The application of artificial intelligence, machine vision, and other sensor technologies in aquaculture enables real-time monitoring of the ecological environment. Combined with deep learning image recognition and detection, as well as machine learning methods, this allows for unified analysis of marine organisms' living habits, feeding intensity, environmental adaptation, and ecological and environmental information. This allows for in-depth analysis of aquaculture process data, improving efficiency and decision-making reliability.

[0004] However, due to the diverse biodiversity and complex biological environment of offshore fishing grounds, as well as the turbid water and significant underwater scattering effects, the difficulty of underwater target recognition is significantly different from that of terrestrial imagery. Under these special conditions, using target detection technology to identify underwater targets may face low accuracy and a high incidence of false detections and missed detections. Summary of the Invention

[0005] The purpose of the present invention is to address the problems of low accuracy and easy false detection and missed detection when using target detection technology to identify underwater targets in current offshore fisheries, and to provide an improved method for high-precision target detection in marine fisheries based on RT-DETR. By balancing computational efficiency and recognition accuracy, and utilizing multi-level feature fusion and the introduction of a small target detection layer, it is possible to better capture multi-level semantic features in underwater remote sensing images and improve target recognition while ensuring network lightweight, ignoring the influence of complex background to improve detection effect, and providing methodological theory and data support for the development of intelligent monitoring equipment for marine fisheries, marine ecological protection, and fishery insurance.

[0006] The present invention provides the following technical solution: an improved method for high-precision target detection in marine fisheries based on RT-DETR, the improved method for high-precision target detection in marine fisheries comprising the following steps:

[0007] S1. Collect and preprocess multi-view images of marine life in offshore fishing grounds to obtain a marine fishing ground target dataset.

[0008] S2. Use the Efficient Block module to replace the BasicBlock module in the backbone network of the RT-DETR model, and introduce a small target detection layer to improve the feature extraction capability of the RT-DETR model;

[0009] S3. Introducing the cascaded grouped convolutional attention module to replace the AIFI module in the RT-DETR model to improve the computational efficiency of the RT-DETR model;

[0010] S4. Introducing a cross-scale fusion module based on the attention mechanism into the neck network of the RT-DETR model to enhance feature flow and strengthen the RT-DETR model's ability to acquire multi-scale information;

[0011] S5. Configure a training environment to train the input improvement method of the marine fishery target data set;

[0012] S6. Apply the trained improved high-precision target detection method, take the target image to be detected as input, and obtain the marine fishery target detection result and marine fishery target recognition information as output.

[0013] As a further technical solution of the present invention: in step S1, the multi-perspective images of marine life in offshore fishing grounds include data received by underwater robot sensors and images taken by divers' manual cameras; preprocessing includes imaging quality screening, data size cropping, marine life identification label production and data set division.

[0014] As a further technical solution of the present invention: imaging quality screening and data size cropping are completed by professional marine science technicians, and the final input data obtained is an RGB image containing three types of marine organisms: sea cucumbers, sea urchins, and scallops; the content of the marine organism identification label production includes the category of the marine organism bounding box, the coordinates of the upper left corner of the bounding box, and the coordinates of the lower right corner of the bounding box; the data set is divided into a training set: validation set: test set ratio of 7:2:1.

[0015] As a further technical solution of the present invention: in step S2, the Basic Block module in the backbone network of the RT-DETR model is replaced by the Efficient Block module, specifically including:

[0016] S21. In the initial stage, in order to minimize the impact of downsampling on feature extraction, the initial embedding layer consists of a 3×3 ConvNorm module with a stride of 1, a 3×3 ConvNorm module with a stride of 2, and a maximum pooling layer. The ConvNorm module refers to the sequential processing of feature maps using convolutional layers, batch normalization layers, and SiLU activation functions.

[0017] In the feature extraction stage, four Efficient Block modules are used to downsample the input feature map. The Efficient Block module consists of a special downsampling residual block and a partial convolution residual block. The special downsampling residual block is combined with the maximum pooling layer and the average pooling layer to build a residual connection network to achieve the purpose of spatial downsampling and channel expansion. The partial convolution residual block consists of a PConv convolution layer and two 1×1 convolution layers to form a residual structure.

[0018] In the feature output stage, four Efficient Block modules are used to downsample the input feature map, helping the RT-DETR model capture the global information in the input data and retain the most important features, thus obtaining four feature maps of different scales: P1, P2, P3, and P4.

[0019] As a further technical solution of the present invention: in step S3, the cascaded grouped convolutional attention module includes a CGA mechanism, specifically including: the CGA mechanism of the cascaded grouped convolutional attention module is inspired by grouped convolution, and the head design of the cascaded grouped convolutional attention module is split. The split processing means that grouping is started before calculating Q, K and V, and Q, K and V are no longer split when calculated internally.

[0020] As a further technical solution of the present invention: in step S4, the attention mechanism includes upper network spatial attention and lower network channel attention;

[0021] The spatial attention of the upper network uses the deformable convolutional layer DCNv2 as a context aggregator for the local channel to extract spatial features; the channel attention of the lower network uses multiple one-dimensional channel convolutions to extract features from adjacent parts of the feature map, specifically including:

[0022] S41, upper network spatial attention processing, upper network spatial attention of global features S att The calculation formula is as follows:

[0023]

[0024] Where: CBR module refers to the sequential processing of feature maps using 1×1 convolution layer, batch normalization layer and ReLU activation function; DBR module extracts spatial features between feature maps of different input paths through deformable convolution; X and Y are feature maps from different input paths; symbol represents the channel dimension serial superposition, then represents the sum of X and Y channel features; B represents the batch normalization layer, and δ represents the ReLU activation function; is a convolutional layer with a convolution kernel of 1×1; DCNv2(·) is a deformable convolutional layer;

[0025] S42, the lower network channel attention processing includes halving the number of input feature channels through 1×1 convolution, then taking the pooled result as input, performing one-dimensional convolution with kernel sizes of 3, 5, and 7, and finally superimposing the channels and using 1×1 convolution to restore the number of channels to be consistent with the original input. The lower network channel attention C of the global feature att The calculation formula is as follows:

[0026]

[0027] Where: Gap(·) represents the global average pooling layer; Represents a one-dimensional convolution with a convolution kernel size of k;

[0028] S43, fusion output, using the broadcast mechanism to sum the processing results of the upper network spatial attention and the processing results of the lower network channel attention, and apply the Sigmoid activation function to perform weighted output. The output result is shown as follows:

[0029]

[0030] Where: Output Z represents the output result of the fusion module; (S att +C att ) represents the summation result of two types of attention features; The symbol represents the element-by-element multiplication of the corresponding elements of the two types of feature maps; w is the fusion weight consisting of a real number between 0 and 1.

[0031] As a further technical solution of the present invention: in step S5, a training environment is configured to train the input improved method of the marine fishery target data set, wherein the training parameters are: 200 training rounds; the input image resolution size is 640×640 pixels; the number of iterative training sample groups is 8 groups; the training process optimizer is the AdamW function; the initial learning rate is 0.0001; and the weight decay is 0.0001.

[0032] As a further technical solution of the present invention: in step S6, the improved method for high-precision target detection to complete training meets the conditions including the maximum number of training rounds, loss function constraints and early stopping mechanism requirements; the marine fishery target detection results and marine fishery target recognition information finally output include biological species, recognition confidence and the position information of the recognized target in the image.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1) Based on the RT-DETR target detection network, the network backbone is redesigned to introduce a small target detection layer. Utilizing a fast feature extraction network structure, this ensures fast computational efficiency while making the network model lightweight for easier deployment and training.

[0035] 2) Improvements to the encoder portion of the RT-DETR object detection network: Firstly, the AIFI module of the RT-DETR object detection network is modified based on the cascaded group convolutional attention module, which greatly reduces the network's redundant computational effort. Secondly, a cross-scale fusion module with an attention mechanism is introduced into the neck network of the RT-DETR model to address the RT-DETR model's limited ability to extract contextual information, enhance feature flow, and strengthen the model's ability to acquire multi-scale information.

[0036] 3) The present invention achieves a good balance between target recognition accuracy and network volume. It improves the positioning accuracy and recognition accuracy in the task of detecting small-volume, large-number, and multi-species marine biological targets in marine fisheries, and can enhance the detection capabilities in complex marine fishery scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A schematic diagram of the overall steps of the present invention;

[0038] Figure 2 A schematic diagram of the overall network structure of the present invention;

[0039] Figure 3 This is a schematic diagram of the structure of the backbone network Efficient Feature Extractor in the present invention;

[0040] Figure 4 Schematic diagram of the structure of the Focus Fusion Module, a cross-scale fusion module based on spatial attention and channel attention in the decoder part of the present invention;

[0041] Figure 5 Schematic diagram of the original image of the test data set in an embodiment of the present invention;

[0042] Figure 6Schematic diagram of the detection results of the RT-DETR target detection model in an embodiment of the present invention;

[0043] Figure 7 This is a schematic diagram of the detection results of the improved method for high-precision target detection in marine fisheries based on RT-DETR in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0045] Example 1, as Figure 1 and Figure 2 As shown, this embodiment provides an improved method for high-precision target detection in marine fisheries based on RT-DETR, and the improved method for high-precision target detection in marine fisheries includes the following steps:

[0046] First: Collect and preprocess multi-perspective images of marine life in offshore fishing grounds to obtain a marine fishing ground target dataset.

[0047] Among them, the multi-perspective images of marine life in offshore fishing grounds include data received by underwater robot sensors and images taken by divers' manual cameras. The resolution of the multi-perspective images is 3840×2160; the preprocessing includes imaging quality screening, data size cropping, marine life identification label production and data set division.

[0048] like Figure 5 The following figure shows multi-perspective images of marine life in a fishing ground in a coastal area (the original images of the dataset). Image quality screening and data size cropping (or dataset acquisition and processing) were completed by marine science professionals. The final input data is an RGB image of three types of marine life: sea cucumbers, sea urchins, and scallops, with a resolution of 640×640. The content of the marine life identification label production includes the category of the marine life bounding box, the coordinates of the upper left corner of the bounding box, and the coordinates of the lower right corner of the bounding box. The dataset is divided into a training set: validation set: test set ratio of 7:2:1.

[0049] Second: Use the Efficient Block module to replace the BasicBlock module in the backbone network of the RT-DETR model, and introduce a small target detection layer to improve the feature extraction capability of the RT-DETR model.

[0050] The small target detection layer, also known as the small target feature layer, has different effects due to the different number of information extraction layers in the RT-DETR model. The small target detection layer is mainly used to detect relatively small targets. Its feature map volume is relatively large and the spatial resolution is high, which can obtain more feature information of small-sized targets. In addition, the structure of the backbone network in the RT-DETR model, such as Figure 3 shown.

[0051] Use the Efficient Block module to replace the Basic Block module in the backbone network of the RT-DETR model. The specific steps include the following:

[0052] 1) In the initial stage, in order to minimize the impact of downsampling on feature extraction, the initial embedding layer consists of a 3×3 ConvNorm module with a stride of 1, a 3×3 ConvNorm module with a stride of 2, and a maximum pooling layer. The ConvNorm module refers to the sequential processing of feature maps using convolutional layers, batch normalization layers, and SiLU activation functions. It should be noted that in order to optimize the traditional convolutional feature extraction process, a 1×1 convolutional layer is used to reduce the number of channels before the downsampling operation;

[0053] 2) In the feature extraction stage, four Efficient Block modules are used to downsample the input feature map. The Efficient Block module consists of a special downsampling residual block and a partial convolution residual block. The special downsampling residual block combines the maximum pooling layer and the average pooling layer to build a residual connection network to achieve the purpose of spatial downsampling and channel expansion. The partial convolution residual block consists of a PConv convolution layer and two 1×1 convolution layers to form a residual structure. In addition, the PConv convolution layer performs convolution operations on a portion of the input feature map instead of applying it to the entire map, significantly reducing redundant computation and memory access.

[0054] In the Efficient Block module, the input features are first segmented to obtain two parts of feature information, named feature information 1 and feature information 2. Feature information 1 is downsampled by a 3×3 ConvNorm layer with a step size of 1 and a 3×3 ConvNorm layer with a step size of 2 to extract features and obtain output information a. Feature information 2 is segmented again to obtain two parts of feature information, named feature information 2-1 and feature information 2-2. Feature information 2-1 is downsampled by a maximum pooling layer and a 1×1 ConvNorm layer with a step size of 1 to obtain output information b, and feature information 2-2 is downsampled by an average pooling layer and a step size of 1. The 1×1 ConvNorm layer is used to downsample the information to obtain output information c; then, through the channel number connection layer, the output information b is connected to the output information c to obtain output information d, and then through the network superposition operation, the output information a and the output information d are superimposed to obtain output information e; through the segmentation operation, the output information e is divided into two parts of feature information, named output information f and output information g respectively, where the output information g is sequentially subjected to a PConv convolution layer and two 1×1 ConvNorm modules with a stride of 1 to obtain output information h; finally, through the network superposition operation, the output information f and the output information h are superimposed to obtain the output feature, and the EfficientBlock module ends;

[0055] 3) In the feature output stage, the input feature map is downsampled through four Efficient Block modules to help the RT-DETR model capture the global information in the input data and retain the most important features, resulting in four feature maps of different scales: P1, P2, P3, and P4.

[0056] Third: Introduce the cascaded group convolutional attention module to replace the AIFI module in the RT-DETR model to improve the computational efficiency of the RT-DETR model.

[0057] In this embodiment, the AIFI module in the RT-DETR model still adopts a multi-head self-attention structure, but in order to improve the performance of the RT-DETR model and reduce the network complexity, the cascaded grouped convolutional attention module proposed in EfficientVIT is introduced, wherein the CGA mechanism of the cascaded grouped convolutional attention module (abbreviated as CGAM) is inspired by grouped convolution, and the head design of the cascaded grouped convolutional attention module is split (i.e., split into groups). The split processing means that grouping is started before calculating Q, K and V, and Q, K and V are no longer calculated separately internally. In the standard multi-head attention of the original RT-DETR model, Q, K and V are obtained using all input feature channels.

[0058] Here we explain Q, K, and V respectively: 1) Query (Q): used to "ask" for information. It indicates the part of the current input feature that needs to be paid attention to among other features, similar to a "question"; 2) Key (K): used to "match" queries. It is a vector associated with each feature, representing the "identifier" or "keyword" of each position; 3) Value (V): represents the "content of the information". It carries the information of the feature and will eventually be weighted and output.

[0059] Calculate similarity between Q and K: In the self-attention mechanism, the similarity between the query (Q) and the key (K) is first calculated. A common approach is to calculate their dot product, and then normalize them through Softmax to obtain the weight of each feature (i.e., attention weight).

[0060] Weighted value (V): The obtained attention weight is multiplied by the value (V) to obtain the weighted information output. The weight of each value will be adjusted according to the degree of match with the query. The higher the match, the greater the weight and the greater the influence of the information.

[0061] In summary, Q, K, and V are three vectors used to calculate attention weights and weighted information respectively: Q determines the "query content" of the current input; K determines whether each position is similar to the query; V carries the information of each position, which will eventually be weighted and summed.

[0062] Fourth: Introducing a cross-scale fusion module based on the attention mechanism into the neck network of the RT-DETR model to solve the problem of limited context information extraction by the RT-DETR model, thereby enhancing feature flow and strengthening the RT-DETR model's ability to obtain multi-scale information, such as Figure 4 shown.

[0063] Among them, the attention mechanism includes the upper network spatial attention and the lower network channel attention, and the upper network and the lower network are channels for feature extraction.

[0064] The upper network spatial attention uses the deformable convolutional layer DCNv2 as a context aggregator for the local channel to extract spatial features. To keep the algorithm lightweight, only the local context needs to be added to the global context of the attention module. The lower network channel attention uses multiple one-dimensional channel convolutions to extract features from adjacent parts of the feature map. The key idea is to achieve channel attention between multi-scale feature maps by setting the size of a specific pooling calculation layer. The steps include:

[0065] 1) Upper network spatial attention processing, upper network spatial attention S of global features att The calculation formula is as follows:

[0066]

[0067] Where: CBR module refers to the sequential processing of feature maps using 1×1 convolution layer, batch normalization layer and ReLU activation function; DBR module extracts spatial features between feature maps of different input paths through deformable convolution; X and Y are feature maps from different input paths, and Y can usually be considered as the feature map with larger receptive field; symbol represents the channel dimension serial superposition, then represents the sum of X and Y channel features; B represents the batch normalization layer, and δ represents the ReLU activation function; is a convolutional layer with a convolution kernel of 1×1; DCNv2(·) is a deformable convolutional layer;

[0068] 2) The lower network channel attention processing includes halving the number of input feature channels through 1×1 convolution, then taking the pooled result as input, performing one-dimensional convolution with kernel sizes of 3, 5, and 7, and finally superimposing the channels and using 1×1 convolution to restore the number of channels to be consistent with the original input. The lower network channel attention C of the global feature att The calculation formula is as follows:

[0069]

[0070] Where: Gap(·) represents the global average pooling layer; Represents a one-dimensional convolution with a convolution kernel size of k;

[0071] Channel attention uses multiple sets of one-dimensional convolutions to test the correlation between features of different scales output by the target detection network backbone;

[0072] 3) Fusion output: Use the broadcast mechanism to sum the processing results of the upper network spatial attention and the lower network channel attention, apply the Sigmoid activation function and perform weighted output. The output result formula is as follows:

[0073]

[0074] Where: Output Z represents the output result of the fusion module; (S att +C att ) represents the summation result of two types of attention features; The symbol represents the element-by-element multiplication of the corresponding elements of the two types of feature maps; w is the fusion weight composed of real numbers between 0 and 1, so (1-w) is also a real number. This allows the fusion module based on the attention mechanism to perform soft selection or weighted averaging in the input data, and realize the fusion of different feature maps by dynamically allocating attention weights to feature maps. While ensuring lightweight, it ignores the influence of complex background and improves the ability of target recognition.

[0075] Fifth: Configure the training environment to train the improved input method of the marine fishery target data set (improved method for high-precision target detection).

[0076] The training parameters are as follows: 200 training rounds; input image resolution of 640×640 pixels; 8 iterative training sample groups; AdamW function as the training process optimizer; initial learning rate of 0.0001; and weight decay of 0.0001.

[0077] Sixth: Apply the improved method of high-precision target detection that has been trained, take the target image to be detected as input, and obtain the marine fishery target detection results and marine fishery target recognition information as output.

[0078] Among them, the improved method for high-precision target detection to complete training meets the conditions including maximum training rounds, loss function constraints and early stopping mechanism requirements; the final output of marine fishery target detection results and marine fishery target recognition information includes biological species, recognition confidence and the position information of the recognized target in the image.

[0079] In the second embodiment, in an actual example of a marine fishery, since the underwater target detection network is a key part of the underwater detection system deployed in the marine fishery, it is of great significance for underwater rapid positioning, automatic monitoring of marine biological growth, intelligent fishing, etc. Therefore, the best way to improve its performance is to train it with real picked-up environment images, and select the enhanced underwater detection dataset EUDD based on real marine fishery images on the high seas.

[0080] In this embodiment, the dataset is obtained from video recordings of two underwater locations about 500 meters away from Zhangzi Island. The video is completed by a robot and a diver along a specific circular route. The video is sampled and cut according to the definition (720P, 1080P and 4K video), shooting angle (straight view, top-down) and terrain scene (such as flat ground, slope, stone), and multiple categories of images are sampled and cut according to a uniform number of frames. The final constructed underwater open sea target detection dataset contains 2227 original images of three major categories (sea cucumbers, sea urchins, and scallops). The original images of the three seabed creatures are shown in the attached figure. Figure 5 shown.

[0081] The training network parameters were set as follows: 200 training rounds; 640×640 pixel input image resolution; 8 iterative training sample groups; AdamW optimizer; 0.0001 initial learning rate; and 0.0001 weight decay. Table 1 lists the hardware and software configurations for the network training platform.

[0082] Table 1 is the network training software and hardware configuration table

[0083] Configuration parameter operating system Ubuntu 16.04 programming language Python 3.9 CPU 12th Gen Intel(R)Core(TM)i7-12700K 3.60GHz GPU GeForce RTX 3090 GPU memory 24G CUDA version 12.0 cuDNN version 11.7 RAM capacity 64G Algorithm framework Pytorch-2.0.1+cu117 Programming software Spyder 3.3.0

[0084] The target image to be detected is used as input, and the marine fishery target detection results and marine fishery target recognition information are obtained as output. The conditions for completing the training of the improved method for high-precision target detection include the maximum training rounds, loss function constraints, and early stopping mechanism requirements; the final output of the marine fishery target detection results includes the species, recognition confidence, and the position information of the recognized target in the image. On the test dataset, the detection results of the improved method for high-precision target detection in marine fisheries based on RT-DETR are as follows: Figure 7 shown.

[0085] Table 2 is the target detection results statistics table

[0086]

[0087] Through comparative experiments with the original RT-DETR, it is found that the detection results of the RT-DETR target detection model on the test dataset are as follows: Figure 6 As shown. Table 2 shows the test results of the improved algorithm proposed in the present invention and the RT-DETR network in the underwater biological detection data set. The input data is the test set data divided by EUDD. Among them, the number of sea urchins detected is greater than the sum of sea cucumber and shell data, which is consistent with the actual species distribution. It can be seen that compared with the RT-DETR network, the improved algorithm proposed in the present invention has higher accuracy in the recognition of the three types of underwater biological targets. However, in the process of sea cucumber target detection, the number of instances and pictures recognized by RT-DETR is greater than that of the improved algorithm proposed in the present invention, and the recall rate is also higher, but the recognition accuracy is far behind that of the improved algorithm proposed in the present invention. In the detection of sea urchin targets with a large number and small size, the improved algorithm proposed in the present invention shows its more accurate advantage, identifying more sea urchin instances in a smaller number of pictures, and its accuracy and recall rate are better than the RT-DETR network. In terms of the detection accuracy of the three underwater biological categories of the two networks, the mAP50 of shells is the highest, which is related to the biological attribute of underwater shell features that are easy to distinguish.

[0088] Based on the recognition effects of three underwater creatures in the EUDD dataset involved in this paper, compared with the RT-DETR network before optimization, the improved algorithm proposed in this paper further improves the accuracy of target recognition while meeting the lightweight requirements of the network. The accuracy is increased from 89.2% to 91.4%; the recall rate is increased from 85.2% to 87.1%; the mAP50 is increased from 88.5% to 92.7%, and the network recognition accuracy is improved by 4.7%. The detailed comparison statistics are shown in Table 3 below.

[0089] Table 3 is the statistical table of comparative experiments

[0090]

[0091] In summary, the improved algorithm proposed in this paper can effectively identify and accurately locate multi-category, large number, and small-volume underwater biological targets in complex underwater environments. The network has contributed to the advancement of underwater target detection tasks and provided a reliable solution for target detection in actual complex ocean scenes.

[0092] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

[0093] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. An improved method for high-precision target detection in marine fisheries based on RT-DETR, characterized in that: The improved method for high-precision target detection in marine fisheries comprises the following steps: S1. Collect and preprocess multi-view images of marine organisms in marine fisheries to obtain a marine fishery target dataset. S2. Use the Efficient Block module to replace the Basic Block module in the backbone network of the RT-DETR model, and introduce a small target detection layer to improve the feature extraction capability of the RT-DETR model. The Efficient Block module consists of a special downsampling residual block and a partial convolution residual block: the special downsampling residual block combines the maximum pooling layer and the average pooling layer to construct a residual connection network to achieve the purpose of spatial downsampling and channel expansion; the partial convolution residual block consists of a PConv convolution layer and two 1×1 convolution layers to form a residual structure; S3. Introducing the cascaded grouped convolutional attention module to replace the AIFI module in the RT-DETR model to improve the computational efficiency of the RT-DETR model; S4. Introducing a cross-scale fusion module based on the attention mechanism into the neck network of the RT-DETR model to enhance feature flow and strengthen the RT-DETR model's ability to acquire multi-scale information. The attention mechanism includes upper network spatial attention and lower network channel attention. The upper network spatial attention uses the deformable convolution layer DCNv2 as a context aggregator for the local channel to extract spatial features; the lower network channel attention uses multiple one-dimensional channel convolutions to extract features from adjacent parts of the feature map. S5. Configuring a training environment to train the input improvement method of the marine fishery target data set; S6. Apply the trained improved high-precision target detection method, take the target image to be detected as input, and obtain the marine fishery target detection result and marine fishery target recognition information as output.

2. The improved method for high-precision target detection in marine fisheries according to claim 1, characterized in that: In step S1, the multi-view images of marine life include data received by underwater robot sensors and images taken by diver's manual camera; the preprocessing includes imaging quality screening, data size cropping, marine life identification label production and data set division.

3. The improved method for high-precision target detection in marine fisheries according to claim 2, characterized in that: The imaging quality screening and data size cropping are completed by professional marine science technicians. The final input data is RGB images containing three types of marine organisms: sea cucumbers, sea urchins, and scallops. The content of the marine organism identification label production includes the category of the marine organism bounding box, the coordinates of the upper left corner of the bounding box, and the coordinates of the lower right corner of the bounding box. The data set is divided into a training set: validation set: test set ratio of 7:2:

1.

4. The improved method for high-precision target detection in marine fisheries according to claim 1, characterized in that: In step S2, the Basic Block module in the backbone network of the RT-DETR model is replaced with the Efficient Block module, which specifically includes: S21. In the initial stage, in order to minimize the impact of downsampling on feature extraction, the initial embedding layer consists of a 3×3 ConvNorm module with a stride of 1, a 3×3 ConvNorm module with a stride of 2, and a maximum pooling layer. The ConvNorm module refers to the sequential processing of feature maps using convolutional layers, batch normalization layers, and SiLU activation functions. S22, feature extraction stage, uses four Efficient Block modules to downsample the input feature map; In the feature output stage, four Efficient Block modules are used to downsample the input feature map, helping the RT-DETR model capture the global information in the input data and retain the most important features, thus obtaining four feature maps of different scales: P1, P2, P3, and P4.

5. The improved method for high-precision target detection in marine fisheries according to claim 1 is characterized in that: In step S3, the cascaded grouped convolutional attention module includes a CGA mechanism, specifically including: the CGA mechanism of the cascaded grouped convolutional attention module is inspired by grouped convolution, and the head design of the cascaded grouped convolutional attention module is split. The split processing means that grouping is started before calculating Q, K and V, and Q, K and V are no longer split when calculated internally.

6. The improved method for high-precision target detection in marine fisheries according to claim 1 is characterized in that: Step S4 specifically includes: S41, upper network spatial attention processing, upper network spatial attention of global features S att The calculation formula is as follows: Where: CBR module refers to the sequential processing of feature maps using 1×1 convolution layer, batch normalization layer and ReLU activation function; DBR module extracts spatial features between feature maps of different input paths through deformable convolution; X and Y are feature maps from different input paths; symbol represents the channel dimension serial superposition, then represents the sum of X and Y channel features; B represents the batch normalization layer, and δ represents the ReLU activation function; is a convolutional layer with a convolution kernel of 1×1; DCNv2(·) is a deformable convolutional layer; S42, the lower network channel attention processing includes halving the number of input feature channels through 1×1 convolution, then taking the pooled result as input, performing one-dimensional convolution with kernel sizes of 3, 5, and 7, and finally superimposing the channels and using 1×1 convolution to restore the number of channels to be consistent with the original input. The lower network channel attention C of the global feature att The calculation formula is as follows: Where: Gap(·) represents the global average pooling layer; Represents a one-dimensional convolution with a convolution kernel size of k; S43, fusion output, using the broadcast mechanism to sum the processing results of the upper network spatial attention and the processing results of the lower network channel attention, and apply the Sigmoid activation function to perform weighted output. The output result is shown as follows: Where: Output Z represents the output result of the fusion module; (S att +C att ) represents the summation result of two types of attention features; The symbol represents the element-by-element multiplication of the corresponding elements of the two types of feature maps; w is the fusion weight consisting of a real number between 0 and 1.

7. The improved method for high-precision target detection in marine fisheries according to claim 1 is characterized in that: In step S5, a training environment is configured to train the input improvement method of the marine fishery target data set, wherein the training parameters are: 200 training rounds; the input image resolution size is 640×640 pixels; the number of iterative training sample groups is 8 groups; the training process optimizer is the AdamW function; the initial learning rate is 0.0001; and the weight decay is 0.0001.

8. The improved method for high-precision target detection in marine fisheries according to claim 1 is characterized in that: In step S6, the improved method for high-precision target detection that completes training meets the conditions including the maximum number of training rounds, loss function constraints and early stopping mechanism requirements; the marine fishery target detection results and marine fishery target recognition information finally output include biological species, recognition confidence and the position information of the recognized target in the image.

Images