Intelligent Target Detection Method for Deep-Sea Multi-Scale Targets Based on Individual-Group Collaborative Modeling

Abstract

This application belongs to the field of intelligent target detection technology in deep-sea scenarios, specifically involving an intelligent multi-scale target detection method in deep-sea environments based on individual-group collaborative modeling. This method improves the robustness of the lightweight model to degraded deep-sea images through two-stage training, and combines individual features and group features for collaborative modeling to reduce the interference of spatial co-occurrence and group distribution on the detection of individual targets. At the same time, it introduces a hierarchical detection strategy for multi-scale targets, taking into account the detection representation of targets at different scales, thereby improving the accuracy and stability of intelligent target detection in complex deep-sea scenarios.

Description

Technical Field

[0001] This application belongs to the field of multi-scale intelligent target detection technology in deep-sea scenarios, specifically involving a multi-scale intelligent target detection method for deep-sea targets based on individual-group collaborative modeling. Background Technology

[0002] Intelligent target detection technology in complex deep-sea scenarios has significant application value in seabed resource surveys, collaborative operations of underwater equipment, and marine ecological monitoring. However, factors such as low illumination and strong scattering in the deep sea easily lead to image quality degradation, and targets such as polymetallic nodules and fish often exhibit dense co-occurrence, group distribution, and significant scale differences. Existing methods mostly rely on image preprocessing to improve imaging quality, which incurs high computational costs and is not conducive to the real-time deployment of deep-sea equipment. At the same time, existing technologies do not adequately consider the interference of target group distribution and multi-scale representation, resulting in limited stability and accuracy of single target detection in dense scenes. Summary of the Invention

[0003] To address the aforementioned issues, this patent proposes a deep-sea multi-scale target intelligent detection method based on individual-group collaborative modeling. The technical solution is as follows: A deep-sea multi-scale target intelligent detection method based on individual-group collaborative modeling includes the following steps: S1. Construct a deep-sea multi-scale intelligent target detection model based on individual and group collaborative modeling. This model uses deep-sea images... For input objects; S2. Construct a two-stage training dataset consisting of original image samples and corresponding degraded image samples; S3. It facilitates the second-stage training of the deep-sea multi-scale target intelligent detection model using the second-stage training dataset in step S2; S4. Utilize the trained deep-sea multi-scale target intelligent detection model to perform intelligent detection on actual input deep-sea images.

[0004] Preferably, the deep-sea multi-scale target intelligent detection model includes constructing a backbone feature extraction network and extracting multi-level features, specifically: the input deep-sea image is denoted as... The lightweight backbone network is input, and features at different semantic levels are extracted layer by layer through multi-level residual blocks to obtain a four-layer backbone feature. : ; in, These represent downsampling modules one through four, which halve the resolution of the input feature map through convolution. These represent residual blocks one through four, respectively. Mainly retain shallow edge information. and Balancing local details with mid-to-high-level semantic information, It provides stronger information about the distribution of the population.

[0005] Preferably, the deep-sea multi-scale target intelligent detection model includes a module for constructing edge contour information and extracting shallow edge contour features. The specific steps are as follows: Step 1: Feature Analysis Perform upsampling and compare with features The features are superimposed at corresponding pixel positions to form a fused feature. ; Step 2: Analyze the fusion features Perform 3×3 standard deviation statistics based on spatial location on each channel to highlight the edge contour features with drastic local changes; Step 3: Input the local standard deviation statistics into a 1×1 convolutional mapping layer and a sigmoid activation function to obtain the weights of the full-channel edge contour information; finally, use the full-channel edge contour information weights to fuse the features. Channel-by-channel recalibration was performed to obtain edge contour features. : ; in, Indicates convolution upsampling; and These represent 1×1 convolution and 3×3 convolution, respectively; This represents the sigmoid activation function; This represents the standard deviation statistics based on a 3×3 window of spatial location; Indicates the weight of the edge contour information of the entire channel; This indicates element-wise multiplication.

[0006] Preferably, the deep-sea multi-scale target intelligent detection model includes a texture feature information construction module to extract texture features, specifically: Step 1: Feature Analysis Perform upsampling and compare with features The features are superimposed at corresponding pixel positions to form a fused feature. Step 2: Analyze the fusion features Perform 3×3 mean statistics based on spatial location on each channel to highlight internal texture features; Step 3: Input the local mean statistical results into a 1×1 convolutional mapping layer and a sigmoid activation function to obtain the weights of the full-channel texture feature information; Step 4: Utilize the weights of full-channel texture feature information to fuse features Perform channel-by-channel recalibration to obtain texture features. : ; in, Indicates convolution upsampling; and These represent 1×1 convolution and 3×3 convolution, respectively; This represents the sigmoid activation function; This represents the mean statistics based on a 3×3 window of spatial location; Indicates the weights of all-channel texture feature information; This indicates element-wise multiplication.

[0007] Preferably, the deep-sea multi-scale target intelligent detection model includes a dual-index semantic enhancement and reconstruction module to achieve feature modulation "first the subject, then the edge", specifically: Step 1: Let the feature to be modulated be X, based on the texture features output by the texture feature information module. Perform the first modulation; Step 2: Analyze the features after the first modulation. Convolution, batch normalization, and ReLU activation are performed to adapt intermediate features; edge contour features output by the edge contour information module are utilized. For the adapted intermediate features A second modulation is performed to further recover the edge contours and local details of the individual target; Step 3: Modulate the features after the second modulation. Residual fusion is performed with the feature X to be modulated to complete dual-cable sequential modulation, resulting in the output feature Y: ; in, and These represent texture feature modulation blocks and edge contour modulation blocks, respectively. ReLU activation is indicated; BN indicates batch normalization. Represents a 1×1 convolution; Preferably, texture feature modulation block and edge contour modulation block The same lightweight modular structure is adopted, specifically: First, the input features are split into two parts by channel splitting. One part is subjected to channel attention gating, and the other part is subjected to spatial attention gating. This achieves the coordinated modulation of spatial and channel dimensions while controlling the amount of computation. Modulation block and The specific structures are as follows: ; ; Here, `split` represents the channel splitting operation, which is applied to the input features. and The features after splitting are as follows: and ; and A weighted graph representing the channel attention gating branches; and A weighted graph representing the spatial attention gating branches; GAP stands for Global Average Pooling; and This indicates the channel attention gating result; and This represents the spatial attention gating result; concat represents a channel concatenation operation.

[0008] Preferably, the deep-sea multi-scale target intelligent detection model includes constructing a scale-adaptive detection head to complete intelligent detection of targets at different scales. Specifically, let the features modulated by the four dual-index semantic enhancement and reconstruction modules be as follows: ,in These represent characteristics from shallow to deep; First, a layered detection strategy for multi-scale targets is implemented, modulating features in the shallower layers. Features aligned with upsampling By fusing the data, small-scale detection features are obtained. , It primarily preserves high-resolution local detail information, making it suitable for detecting fish at long distances, small multimetallic nodules, and densely packed small targets. ; Mid-layer modulation features Features aligned with upsampling By fusing the data, mesoscale detection features are obtained. : ; Deeper modulation features Features aligned with upsampling By fusing the data, large-scale detection features can be obtained. , It has stronger overall semantic expression capabilities and is suitable for detecting large targets in the near field and large-scale clustered targets: ; Subsequently, the fused multi-scale detection features are input into the corresponding detection heads, and each detection head outputs the bounding box regression results and category prediction results at the corresponding scale. Finally, the detection results at each scale are decoded to the original image coordinate system, and duplicate detection boxes are removed by non-maximum suppression, retaining the target boxes with high confidence, thus completing the target detection.

[0009] Preferably, a two-stage training dataset is constructed, specifically as follows: S21. Obtain raw image data from complex deep-sea scenes, perform sample evaluation and labeling processing on the raw visual data, and generate a training sample set containing deep-sea target categories and bounding box information; wherein, the labeling information includes target category labels and corresponding bounding box position labels; S22. Construct a degradation enhancement sample set based on the original image: Apply a perturbation to the original image that matches the degradation features of deep-sea imaging to generate a set that is similar to the original image. Figure 1 A corresponding degraded image sample; S23. Construct a two-stage training dataset consisting of original image samples and corresponding degraded image samples.

[0010] Preferably, the first stage training of the deep-sea multi-scale target intelligent detection model is used to enhance the robust representation ability of the backbone feature extraction network for degraded deep-sea images, specifically as follows: Original image Corresponding degraded image Each input is fed into the initial training network, and features are extracted using a feature extractor with shared parameters. Based on the detection task, the corresponding prediction results are output. and ; In the first phase of training, the original annotations were used. and To ensure consistency between the prediction results of the original image and the degraded image, a supervision-based consistency loss is introduced. The total loss in the first stage can be expressed as: ; in, y represents the detection supervision loss, which includes classification loss and bounding box loss; b represents the class label of each target; and b represents the ground truth bounding box label of the target. Indicates consistency monitoring loss; This represents the consistency constraint weight.

[0011] Preferably, after completing the first stage of pre-training, the backbone feature extraction network parameters learned in the first stage are loaded into the complete deep-sea multi-scale target intelligent detection model, and then the model enters the second stage of joint optimization training. The loss function used in this stage is: ; in, This indicates the prediction results of the second-stage model.

[0012] Compared with the prior art, the beneficial effects of this application are as follows: 1. By introducing edge contour information module, texture feature information module, and dual-index semantic enhancement and reconstruction module, collaborative modeling of individual target features and group distribution interference is realized, which effectively reduces the problems of false detection, false detection and unstable category discrimination under the dense co-occurrence conditions of targets such as fish and polymetallic nodules in deep-sea scenes.

[0013] 2. A multi-scale hierarchical detection mechanism was constructed, which extracts corresponding detection features for small-scale, medium-scale and large-scale targets and performs detection, thereby realizing targeted modeling and joint detection of deep-sea targets at different scales.

[0014] 3. A two-stage training mechanism was constructed, which improved the robust feature extraction capability of the lightweight detection model for degraded deep-sea images without relying on additional complex image preprocessing procedures. Attached Figure Description

[0015] Figure 1 This is a model architecture diagram; Figure 2 (a) Dual-index semantic enhancement and reconstruction module, (b) Texture feature modulation block, (c) Edge contour modulation block. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] A deep-sea multi-scale target intelligent detection method based on individual-group collaborative modeling includes a stage of constructing a deep-sea multi-scale target intelligent detection model based on individual-group collaborative modeling, a two-stage training dataset construction stage, a two-stage model training stage, and a deep-sea target intelligent detection stage.

[0018] Step 1: Construct a deep-sea multi-scale intelligent target detection model based on individual and group collaborative modeling. This model uses deep-sea images... As the input object, the method first extracts feature information at different levels through the backbone network. During the feature extraction stage, an edge contour information module and a texture feature information module are introduced to recover individual target features from shallow edge contour cues and mid-level internal texture cues, respectively. Furthermore, a dual-index semantic enhancement and reconstruction module is constructed to perform lightweight modulation on the features in a "subject first, edge later" order, thereby reducing the interference of group distribution and spatial co-occurrence on individual target detection. Combined with a scale-adaptive detection head, stable recognition of targets at different scales is achieved. The specific steps of this method are as follows: ① Construct a backbone feature extraction network and extract multi-level features. Let the input deep-sea image be denoted as... The lightweight backbone network is input, and features at different semantic levels are extracted layer by layer through multi-level residual blocks to obtain a four-layer backbone feature. : ; in, These represent the downsampling modules, which halve the resolution of the input feature map through convolution. These represent residual blocks one through four, respectively.

[0019] Mainly retain shallow edge information. and Balancing local details with mid-to-high-level semantic information, This step provides stronger information about the group distribution. Through this step, a multi-level feature base is obtained, necessary for subsequent individual feature preservation and sequential modulation.

[0020] ② Construct an edge contour information module to extract shallow edge contour features. To avoid the shallow individual edge cues being submerged in group distribution scenarios, an edge contour information module is constructed, and standard deviation statistics are used to enhance the response of individual boundary regions. The specific steps are as follows: First, regarding features Perform upsampling and compare with features The features are superimposed at corresponding pixel positions to form a fused feature. Then, the fusion features were analyzed. A 3×3 standard deviation statistical analysis based on spatial location is performed on each channel to highlight edge contour features with drastic local changes. The results of this local standard deviation analysis are then input into a 1×1 convolutional mapping layer and a sigmoid activation function to obtain edge contour information weights for all channels. Finally, these weights are used to refine the fused features. Channel-by-channel recalibration was performed to obtain edge contour features. : ; in, Indicates convolution upsampling; and These represent 1×1 convolution and 3×3 convolution, respectively; This represents the sigmoid activation function; This represents the standard deviation statistics based on a 3×3 window of spatial location; Indicates the weight of the edge contour information of the entire channel; This indicates element-wise multiplication.

[0021] ③ Construct a texture feature information module and extract texture features. To enhance the stable representation of the main region within a single target and suppress the misleading co-occurrence caused by the dense distribution of neighboring targets, a texture feature information module is constructed, using mean statistics to highlight internal texture features. The specific steps are as follows: First, regarding features Perform upsampling and compare with features The features are superimposed at corresponding pixel positions to form a fused feature. Then, the fusion features were analyzed. 3×3 mean statistics based on spatial location are performed on each channel to highlight internal texture features. The local mean statistics results are then input into a 1×1 convolutional layer and a sigmoid activation function to obtain full-channel texture feature weights. Finally, these full-channel texture feature weights are used to fuse the features. Perform channel-by-channel recalibration to obtain texture features. : ; in, Indicates convolution upsampling; and These represent 1×1 convolution and 3×3 convolution, respectively; This represents the sigmoid activation function; This represents the mean statistics based on a 3×3 window of spatial location; Indicates the weights of all-channel texture feature information; This indicates element-wise multiplication.

[0022] ④ Construct a dual-index semantic enhancement and reconstruction module to achieve "subject-first, edge-later" feature modulation. To further inject the above two types of individual cues into the main features and achieve co-occurrence interference suppression in a lightweight manner, a dual-index semantic enhancement and reconstruction module is constructed. The specific steps are as follows: Let the feature to be modulated be X. This module first uses the texture features output by the texture feature information module. Perform the first modulation; then analyze the features after the first modulation. Convolution, batch normalization, and ReLU activation are performed to adapt intermediate features; edge contour features output by the edge contour information module are utilized. For the adapted intermediate features A second modulation is performed to further recover the edge contours and local details of the individual target; finally, the features after the second modulation are... The residual fusion is performed with the feature X to be modulated to complete the dual-clue sequential modulation and obtain the output feature Y.

[0023] ; in, and These represent texture feature modulation blocks and edge contour modulation blocks, respectively. ReLU activation is indicated; BN indicates batch normalization. Represents a 1×1 convolution; Among them, texture feature modulation block and edge contour modulation block The same lightweight modular structure is adopted. First, the input features are split into two parts through channel splitting. One part undergoes channel attention gating, and the other part undergoes spatial attention gating, achieving coordinated modulation of the spatial and channel dimensions while controlling computational load. Modulation block and The specific structures are as follows: ; ; Here, `split` represents the channel splitting operation, which is applied to the input features. and The features after splitting are as follows: and ; and A weighted graph representing the channel attention gating branches; and The weight graph represents the spatial attention gating branch; GAP represents global average pooling. and This indicates the channel attention gating result; and This indicates the spatial attention gating result; concat indicates the channel splicing operation.

[0024] ⑤ Construct a scale-adaptive detection head to achieve intelligent detection of targets at different scales. To address the issues of significant differences in target scale and dense co-occurrence in complex deep-sea scenarios, a scale-adaptive detection head is constructed based on multi-level features processed by the dual-index semantic enhancement and reconstruction module. This head performs joint classification and localization detection on small-scale, medium-scale, and large-scale targets, respectively. The specific steps are as follows: Let the features modulated by the four dual-index semantic enhancement and reconstruction modules be as follows: ,in These represent features from shallow to deep. First, a layered detection strategy for multi-scale targets is implemented, modulating features from the shallower layers. Features aligned with upsampling By fusing the data, small-scale detection features are obtained. , It primarily preserves high-resolution local detail information, making it suitable for detecting fish at long distances, small multimetallic nodules, and densely packed small targets. ; Mid-layer modulation features Features aligned with upsampling By fusing the data, mesoscale detection features are obtained. : ; Deeper modulation features Features aligned with upsampling By fusing the data, large-scale detection features can be obtained. , It has stronger overall semantic expression capabilities and is suitable for detecting large targets in the near field and large-scale clustered targets: ; Subsequently, the fused multi-scale detection features are input into the corresponding detection heads, and each detection head outputs the bounding box regression results and category prediction results at the corresponding scale. Finally, the detection results at each scale are decoded to the original image coordinate system, and duplicate detection boxes are removed by non-maximum suppression, retaining the target boxes with high confidence, thus completing the target detection.

[0025] Step 2: Construct the two-stage training dataset.

[0026] First, raw image data of complex deep-sea scenes is acquired. The raw images are then filtered and labeled to obtain a basic training sample set. The labeling information includes target category labels and corresponding bounding box location annotations. Then, based on the original high-quality images, a degradation enhancement sample set is constructed. Specifically, artificial perturbations (such as brightness reduction and contrast attenuation) matching the degradation characteristics of deep-sea imaging are applied to the original images to generate samples that are similar to the original images. Figure 1 One corresponding degraded image sample. Finally, this stage forms a two-stage training dataset consisting of the original high-quality image samples and the corresponding degraded image samples.

[0027] Step 3: Two-stage model training. This stage is used to train the deep-sea multi-scale target intelligent detection model in two stages to improve the model's robustness in complex deep-sea degraded images and enhance its adaptability to target co-occurrence interference, multi-scale distribution, and complex backgrounds.

[0028] ① The first stage of training is used to enhance the robust representation ability of the backbone feature extraction network for degraded deep-sea images. Specifically, the original image... Corresponding degraded image Each input is fed into the initial training network, and features are extracted using a feature extractor with shared parameters. Based on the detection task, the corresponding prediction results are output. and .

[0029] In the first phase of training, on the one hand, the original annotations were used to... and Supervision is implemented to constrain the model to maintain basic detection capabilities. On the other hand, a consistency supervision loss is introduced to ensure that the prediction results for the original and degraded images remain consistent, improving the robustness of the backbone network's feature extraction under changing conditions. The total loss in the first stage can be expressed as: ; in, y represents the detection supervision loss, which includes classification loss and bounding box loss; b represents the class label of each target; and b represents the ground truth bounding box label of the target. Indicates consistency monitoring loss; This represents the consistency constraint weight.

[0030] ② After completing the first stage of pre-training, the backbone feature extraction network parameters learned in the first stage are loaded into the complete deep-sea multi-scale target intelligent detection model, and the second stage of joint optimization training begins. In the second stage, a lower learning rate is used to fine-tune the pre-trained backbone network, while subsequent modules are trained with a normal learning rate.

[0031] The loss function used in this stage is: ; in, This indicates the prediction results of the second-stage model.

[0032] Through the second stage of training, the model, while maintaining the robustness against degradation in the first stage, further gains the ability to adapt to spatial co-occurrence interference, multi-scale distribution, and background complexity in complex deep-sea scenarios, thus forming a complete lightweight intelligent detection model for multi-scale targets in the deep sea.

[0033] Step 4: Use the trained deep-sea multi-scale target intelligent detection model to perform intelligent detection on the actual input deep-sea images.

[0034] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A deep-sea multi-scale target intelligent detection method based on individual-group collaborative modeling, characterized in that, Includes the following steps: S1. Construct a deep-sea multi-scale intelligent target detection model based on individual and group collaborative modeling. This model uses deep-sea images... The input object is first extracted through the backbone network, and then edge contour information module and texture feature information module are introduced in the feature extraction stage to recover the features of individual targets from shallow edge contour cues and middle internal texture cues, respectively. Then, a dual-index semantic enhancement and reconstruction module is constructed to perform lightweight modulation on the features in the order of "subject first, then edge", thereby reducing the interference of group distribution and spatial co-occurrence on the detection of individual targets, and combining with the scale-adaptive detection head to achieve stable recognition of targets at different scales. The deep-sea multi-scale target intelligent detection model includes a dual-index semantic enhancement and reconstruction module to achieve feature modulation "first the subject, then the edge," specifically: Step 1: Let the feature to be modulated be X, based on the texture features output by the texture feature information module. Perform the first modulation; Step 2: Analyze the features after the first modulation. Convolution, batch normalization, and ReLU activation are performed to adapt intermediate features; Edge contour features output by the edge contour information module For the adapted intermediate features Perform a second modulation; Step 3: Modulate the features after the second modulation. Residual fusion is performed with the feature X to be modulated to complete dual-cable sequential modulation, resulting in the output feature Y: ； in, and These represent texture feature modulation blocks and edge contour modulation blocks, respectively. ReLU activation is indicated; BN indicates batch normalization. Represents a 1×1 convolution; S2. Construct a two-stage training dataset consisting of original image samples and corresponding degraded image samples; S3. It facilitates the second-stage training of the deep-sea multi-scale target intelligent detection model using the second-stage training dataset in step S2; S4. Utilize the trained deep-sea multi-scale target intelligent detection model to perform intelligent detection on actual input deep-sea images.

2. The intelligent detection method for deep-sea multi-scale targets based on individual-group collaborative modeling according to claim 1, characterized in that, The deep-sea multi-scale target intelligent detection model includes constructing a backbone feature extraction network and extracting multi-level features. Specifically, the input deep-sea image is denoted as... The lightweight backbone network is input, and features at different semantic levels are extracted layer by layer through multi-level residual blocks to obtain four layers of backbone features. : ； in, These represent downsampling modules one through four, which halve the resolution of the input feature map through convolution. These represent residual blocks one through four, respectively. Mainly retain shallow edge information. and Balancing local details with mid-to-high-level semantic information, It provides stronger information about the distribution of the population.

3. The intelligent detection method for deep-sea multi-scale targets based on individual-group collaborative modeling according to claim 1, characterized in that, The deep-sea multi-scale target intelligent detection model includes a module for constructing edge contour information and extracting shallow edge contour features. The specific steps are as follows: Step 1: Feature Analysis Perform upsampling and compare with features The features are superimposed at corresponding pixel positions to form a fused feature. ; Step 2: Analyze the fusion features Perform 3×3 standard deviation statistics based on spatial location for each channel; Step 3: Input the local standard deviation statistics into a 1×1 convolutional mapping layer and a sigmoid activation function to obtain the weights of the full-channel edge contour information; finally, use the full-channel edge contour information weights to fuse the features. Channel-by-channel recalibration was performed to obtain edge contour features. : ； in, Indicates convolution upsampling; and These represent 1×1 convolution and 3×3 convolution, respectively; This represents the sigmoid activation function; This represents the standard deviation statistics based on a 3×3 window of spatial location; Indicates the weight of the edge contour information of the entire channel; This indicates element-wise multiplication.

4. The intelligent detection method for deep-sea multi-scale targets based on individual-group collaborative modeling according to claim 1, characterized in that, The deep-sea multi-scale target intelligent detection model includes a texture feature information construction module to extract texture features, specifically: Step 1: Feature Analysis Perform upsampling and compare with features The features are superimposed at corresponding pixel positions to form a fused feature. Step 2: Analyze the fusion features Perform 3×3 mean statistics based on spatial location on each channel; Step 3: Input the local mean statistical results into a 1×1 convolutional mapping layer and a sigmoid activation function to obtain the weights of the full-channel texture feature information; Step 4: Utilize the weights of full-channel texture feature information to fuse features Perform channel-by-channel recalibration to obtain texture features. : ； in, Indicates convolution upsampling; and These represent 1×1 convolution and 3×3 convolution, respectively; This represents the sigmoid activation function; This represents the mean statistics based on a 3×3 window of spatial location; Indicates the weights of all-channel texture feature information; This indicates element-wise multiplication.

5. The intelligent detection method for deep-sea multi-scale targets based on individual-group collaborative modeling according to claim 1, characterized in that, Texture feature modulation block and edge contour modulation block The same lightweight modular structure is adopted, specifically: First, the input features are split into two parts by channel splitting. One part is subjected to channel attention gating, and the other part is subjected to spatial attention gating. This achieves the coordinated modulation of spatial and channel dimensions while controlling the amount of computation. Modulation block and The specific structures are as follows: ； ； Here, `split` represents the channel splitting operation, which is applied to the input features. and The features after splitting are as follows: and ; and A weighted graph representing the channel attention gating branches; and A weighted graph representing the spatial attention gating branches; GAP stands for Global Average Pooling; and This indicates the channel attention gating result; and This represents the spatial attention gating result; concat represents a channel concatenation operation.

6. The intelligent detection method for deep-sea multi-scale targets based on individual-group collaborative modeling according to claim 1, characterized in that, The deep-sea multi-scale target intelligent detection model includes constructing a scale-adaptive detection head to complete intelligent detection of targets at different scales. Specifically, the features modulated by four dual-index semantic enhancement and reconstruction modules are as follows: ,in These represent characteristics from shallow to deep; First, a layered detection strategy for multi-scale targets is implemented, modulating features in the shallower layers. Features aligned with upsampling By fusing the data, small-scale detection features are obtained. , Mainly preserves high-resolution local detail information: ； Mid-layer modulation features Features aligned with upsampling By fusing the data, mesoscale detection features are obtained. : ； Deeper modulation features Features aligned with upsampling By fusing the data, large-scale detection features can be obtained. , It has a stronger overall semantic expression capability: ； Subsequently, the fused multi-scale detection features are input into the corresponding detection heads, and each detection head outputs the bounding box regression results and category prediction results at the corresponding scale. Finally, the detection results at each scale are decoded to the original image coordinate system, and duplicate detection boxes are removed by non-maximum suppression, retaining the target boxes with high confidence, thus completing the target detection.

7. The intelligent detection method for deep-sea multi-scale targets based on individual-group collaborative modeling according to claim 1, characterized in that, The two-stage training dataset is constructed as follows: S21. Acquire raw image data in complex deep-sea scenes, perform sample evaluation and labeling processing on the raw image data, and generate a training sample set containing deep-sea target categories and bounding box information; wherein, the labeling information includes target category labels and corresponding bounding box position labels; S22. Based on the original image, construct a degradation enhancement sample set: apply a perturbation to the original image that matches the degradation characteristics of deep-sea imaging to generate degradation image samples that correspond one-to-one with the original image; S23. Construct a two-stage training dataset consisting of original image samples and corresponding degraded image samples.

8. The intelligent detection method for deep-sea multi-scale targets based on individual-group collaborative modeling according to claim 1, characterized in that, The first stage of training for the deep-sea multi-scale target intelligent detection model is used to enhance the robust representation ability of the backbone feature extraction network for degraded deep-sea images, specifically: Original image Corresponding degraded image Each input is fed into the initial training network, and features are extracted using a feature extractor with shared parameters. Based on the detection task, the corresponding prediction results are output. and ; In the first phase of training, the original annotations were used. and To ensure consistency between the prediction results of the original image and the degraded image, a supervision-based consistency loss is introduced. The total loss in the first stage can be expressed as: ； in, y represents the detection supervision loss, which includes classification loss and bounding box loss; b represents the class label of each target; and b represents the ground truth bounding box label of the target. Indicates consistency monitoring loss; This represents the consistency constraint weight.

9. The intelligent detection method for deep-sea multi-scale targets based on individual-group collaborative modeling according to claim 8, characterized in that, After completing the first stage of pre-training, the backbone feature extraction network parameters learned in the first stage are loaded into the complete deep-sea multi-scale target intelligent detection model, and then the model enters the second stage of joint optimization training. The loss function used in this stage is: ； in, This indicates the prediction results of the second-stage model.

Images