Image segmentation system and method for drone seaweed bed monitoring

Abstract

The present invention provides an image segmentation system and method for unmanned aerial vehicle (UAV) seaweed bed monitoring, which belongs to the technical field of image segmentation in computer vision. The system is improved based on an iterative small sample adapter (IFA), and includes a feature grouping mapping part, a feature fusion part, and a segmentation prediction part. The feature grouping mapping part includes a decoder and a grouping mapping module driven by a Gaussian energy prototype, which is used for feature extraction and strengthening attention to key features. The feature fusion part adopts a channel-space feature fusion module to fuse the features extracted by the feature grouping mapping part in channel and spatial dimensions. The grouping mapping module includes a channel attention unit and a spatial attention unit. The segmentation prediction part includes a decoder, which is responsible for converting the optimized feature map into a final segmentation mask. The present invention significantly improves the segmentation accuracy and generalization ability of the model in small sample segmentation tasks through innovative encoders, grouping mapping modules, channel-space feature fusion modules, and decoders.

Description

Technical Field

[0001] The present invention belongs to the technical field of image segmentation based on computer vision, and in particular relates to an image segmentation system and method for unmanned aerial vehicle (UAV) seaweed bed monitoring. Background Art

[0002] With the global degradation of marine ecosystems and the gradual decline of kelp beds, monitoring and protecting them has become increasingly important. Seaweed beds are a key component of marine ecosystems, providing habitat and food sources for numerous marine organisms. Seaweed beds not only play a vital role in maintaining marine biodiversity but also play a crucial role in carbon cycling, climate regulation, and marine ecological balance. However, the distribution and health of seaweed beds are facing serious threats due to factors such as overfishing, pollution, and climate change. Studies have shown that the global coverage of kelp beds has decreased significantly over the past few decades, resulting in the loss of important habitats and food sources for many marine organisms. Traditional methods for monitoring seaweed beds rely primarily on manual diving surveys and underwater videography. These methods are not only time-consuming and labor-intensive, but also have limited coverage, making them inadequate for large-scale monitoring. Manual diving surveys require specialized personnel to conduct underwater operations, which is not only costly but also significantly affected by weather and sea conditions, making them difficult to conduct in harsh environments. While underwater videography can provide relatively intuitive image data, its limited coverage makes it difficult to rapidly monitor large areas of seaweed beds. In addition, traditional methods have large subjectivity and errors in the data acquisition and analysis process, making it difficult to achieve high-precision seaweed bed segmentation and identification.

[0003] The rapid development of drone technology has provided a new solution for seaweed bed monitoring. Drones can capture high-resolution aerial images covering large ocean areas, providing comprehensive data on the distribution and health of seaweed beds. Drone images not only offer high resolution but also enable rapid acquisition of data over large areas, significantly improving monitoring efficiency. Furthermore, drone technology is relatively low-cost, easy to operate, and capable of operating in a variety of weather and sea conditions, offering high flexibility and adaptability. However, image segmentation and recognition of seaweed beds also face numerous challenges. First, the morphology and color of seaweed beds vary significantly across different ocean areas and seasons, making traditional image processing methods ineffective for segmentation. The morphology and color of seaweed beds are influenced by numerous factors, such as water quality, lighting, and temperature, resulting in significant variations across different ocean areas and seasons. Second, drone images often contain complex background information, such as the seawater, sandy bottom, and other marine life, which can interfere with seaweed bed segmentation. The low contrast between seaweed beds and their surroundings makes accurate segmentation particularly challenging against complex backgrounds. Furthermore, image data of kelp beds is often scarce, especially in remote areas, making it difficult to obtain large amounts of labeled data to train deep learning models. This data scarcity limits the model's generalization and segmentation accuracy, restricting its application in diverse waters and under varying conditions.

[0004] Existing small-sample segmentation methods have limitations when processing seaweed bed images. For one thing, these methods typically rely on generating prototype features to match target objects in the query image. However, due to the diversity and complexity of target objects in seaweed bed images, a single prototype feature cannot accurately capture all seaweed bed features, resulting in decreased segmentation accuracy. Furthermore, existing methods often fail to effectively address the differences in feature distribution between the source and target domains when processing cross-domain data, resulting in decreased segmentation accuracy. Seaweed bed images from different ocean regions exhibit significant differences in feature distribution, making it difficult for existing methods to maintain high segmentation accuracy in cross-domain tasks. Summary of the Invention

[0005] To address these challenges, this paper proposes an image segmentation system and method for drone-based seaweed bed monitoring. This system improves segmentation performance by employing a Gaussian Energy Prototype-driven Grouping and Mapping Module (GEPGMM) and a Channel-Spatial Feature Fusion Module (CSFFM). The GEPGMM effectively captures the diversity and complexity of seaweed beds, generating more representative prototype features. Furthermore, the CSFFM utilizes channel- and spatial-attention mechanisms to further enhance feature representation and significantly improve segmentation accuracy, enabling high-precision seaweed bed image segmentation in small-sample segmentation tasks.

[0006] The first aspect of the present invention provides an image segmentation system for UAV seaweed bed monitoring, which is improved based on the iterative small sample adapter (IFA) and takes the seaweed bed image taken by the UAV as input; it includes a feature grouping mapping part, a feature fusion part and a segmentation prediction part;

[0007] The feature group mapping part includes a decoder and a Gaussian energy prototype-driven group mapping module (GEPGMM), which is used to extract features and strengthen the focus on key features. The decoder uses the ResNet-50 model, removing the fully connected layer and the last residual block of the original network, and using the last three network layers of different depths to extract feature information from the input image.

[0008] The feature fusion part adopts the channel space feature fusion module CSFFM to fuse the features extracted by the feature grouping mapping part in the channel and spatial dimensions. The CSFFM module includes a channel attention unit and a spatial attention unit, wherein the channel attention unit generates channel weights through global pooling operation and convolution, and the spatial attention unit generates spatial weights through convolution operation; the channel weights and spatial weights are added together to weight the input feature map to obtain the optimized feature map;

[0009] The segmentation prediction part includes a decoder, which uses three segmentation heads and is responsible for converting the optimized feature map into the final segmentation mask.

[0010] Preferably, the group mapping module GEPGMM includes a seed point selection unit and a feature vector extraction unit. The seed point selection unit first uses the Canny edge detection algorithm to extract the edge features of the input image, and then selects representative seed points by random sampling; the feature vector extraction unit extracts the corresponding feature vector for each seed point, and then calculates the feature distance and spatial distance, and updates the feature point center in combination with the Gaussian energy function to finally generate a multi-scale prototype.

[0011] Preferably, the channel space feature fusion module CSFFM, for the input feature map, where the number of channels, height and width of the feature map are C, H and W respectively, the channel attention module obtains the aggregated features through global average pooling and global maximum pooling, and then uses shared 1D convolution to generate channel weights ; The spatial attention module generates spatial weights through 3×3 convolution ,Finally, the channel weight and spatial weight are added together to weight the input feature map to obtain the optimized feature map.

[0012] Preferably, the decoder adopts the ResNet-50 model, removes the fully connected layer and the last residual block Stage5, and retains the first four stages Stage1-Stage4; the resolution of the feature map output by Stage1 is about 1 / 4 of the input image, containing rich detail information; the resolution of the feature map output by Stage2 is about 1 / 8 of the input image, containing medium-scale features; the resolution of the feature map output by Stage3 is about 1 / 16 of the input image, containing deeper semantic information; the resolution of the feature map output by Stage4 is about 1 / 32 of the input image, containing the deepest semantic information; then, the feature extraction process is performed, and the normalized input image is fed into the ResNet- 50 model, the input image first passes through the first convolutional layer and the maximum pooling layer, the image size is halved, and the shallow detail features are extracted to obtain the feature map of Stage1. Then, the middle-level features are further extracted through the second residual block, and the resolution is halved again to obtain the feature map of Stage2. Then, the deeper semantic features are extracted through the third residual block, and the resolution is halved again to obtain the feature map of Stage3. Finally, the deepest semantic features are extracted through the fourth residual block, and the resolution is further halved to obtain the feature map of Stage4. Finally, the feature grouping and mapping part outputs the feature map; the appropriate feature map is selected from Stage1-Stage4 as the input of the grouping and mapping module GEPGMM.

[0013] Preferably, the specific data processing process of the group mapping module GEPGMM is:

[0014] S1, use the Canny edge detection algorithm to extract the edge features of the support image, generate an edge map to mark the location of the edge points, and randomly sample representative seed points from the edge points as the starting point for subsequent feature grouping;

[0015] S2, for each selected seed point, extract the corresponding feature vector from the feature map Fs of the support image. The feature vector is obtained from the feature map extracted from a certain stage of the ResNet-50 model;

[0016] S3, calculate the feature distance Dfc and spatial distance Dsc between each pixel in the support image feature map Fs and the seed point; the feature distance Dfc is measured by cosine similarity, while the spatial distance Dsc is calculated by Euclidean distance, combined with the smoothing term and energy function, to update the feature point center;

[0017] S4, through the iterative optimization process, gradually updates the feature point centers, and finally generates 5 sets of multi-scale prototypes, which can cover target features of different sizes;

[0018] S5, expands and maps the generated prototypes into the query image for segmentation, and dynamically selects the most appropriate prototype by calculating the similarity between the expanded feature points and the query features;

[0019] S6, generate the response map respmap and the guide map guidemap; the response map respmap represents the maximum similarity between each pixel in the query image and the prototype, while the guide map guidemap represents the prototype index corresponding to each pixel. These two maps are spliced ​​with the original query feature map Fq to generate the optimized query feature a2.

[0020] Preferably, the specific data processing process of the channel space feature fusion module CSFFM is:

[0021] First, the channel attention mechanism is used to enhance the features of important channels in the feature map while suppressing unimportant channels. This is achieved through global average pooling (GAP) and global maximum pooling (GMP), respectively, to obtain two feature vectors A and M, whose shapes are (C, 1, 1), where C is the number of channels in the feature map. These two vectors are then concatenated and passed to a 1×1 convolutional layer to generate channel attention weights Wc. The Sigmoid activation function σ is used to normalize the weights to the range of (0, 1) to highlight important channels and suppress unimportant channels.

[0022] Next, the spatial attention mechanism is used to enhance the features of important spatial locations in the feature map while suppressing unimportant locations. First, the mean μ and maximum max of the input feature map FGEPGMM in the channel dimension are calculated, and its shape is (1, H, W), where H and W are the height and width of the feature map respectively. Then, μ and max are concatenated and passed to a 3×3 convolutional layer to generate the spatial attention weight Ws, which is also normalized using the Sigmoid activation function σ.

[0023] Finally, the channel attention weight Wc and the spatial attention weight Ws are added to obtain the final combined attention weight, and the input feature map FGEPGMM is weighted to generate the optimized feature map F′.

[0024] A second aspect of the present invention provides an image segmentation method for drone seaweed bed monitoring, comprising the following steps:

[0025] Obtain image data of seaweed beds in target areas through drone photography;

[0026] Inputting image data into the image segmentation system according to the first aspect to segment the target area of ​​the input image;

[0027] By performing real-time online analysis on image data and outputting segmentation results, the target area can be accurately segmented.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] In the cross-domain small sample segmentation task, the present invention effectively solves the problems of feature loss and domain mismatch in traditional methods through dynamic prototype grouping and two-dimensional feature fusion, and significantly improves the segmentation accuracy and generalization ability of the model.

[0030] The GEPGMM module uses Gaussian energy for similarity measurement, dynamically generating multiple prototypes based on the feature distribution of the support image. These prototypes not only better capture the characteristics of foreground objects but also effectively avoid the feature loss problem caused by a single prototype in traditional methods. By dynamically selecting the most appropriate prototype, the model can more accurately match foreground objects in the query image, thereby improving segmentation accuracy. In drone-viewed seaweed bed images, this method effectively accounts for the diversity of seaweed bed morphology and color, generating more representative prototypes and improving segmentation accuracy.

[0031] The CSFFM module fuses the features generated by the GEPGMM using both channel and spatial attention mechanisms. The channel attention mechanism enables the model to focus on the most important channel features, while the spatial attention mechanism allows the model to focus on key spatial locations within the image. This dual-dimensional feature fusion not only enhances feature representation capabilities but also effectively preserves and enhances important information in the feature map, further improving segmentation accuracy. In drone-captured seaweed bed images, this approach effectively captures subtle features of the seaweed bed, enhancing segmentation accuracy.

[0032] Through dynamic prototype grouping and two-dimensional feature fusion, the method not only performs well on specific datasets but also maintains high segmentation accuracy across datasets from diverse fields and under varying conditions, demonstrating its strong generalization capabilities and adaptability to a variety of complex practical application scenarios. In drone-captured images of kelp beds, this method can adapt to diverse sea areas, lighting conditions, and morphological variations, demonstrating its strong generalization capabilities and practical application value.

[0033] This paper, through innovative dynamic prototype grouping and two-dimensional feature fusion methods, effectively addresses the feature loss and domain mismatch issues in traditional cross-domain small-sample segmentation methods, significantly improving the segmentation accuracy and generalization capabilities of the model, providing an efficient and practical solution for cross-domain small-sample segmentation tasks. In the segmentation of seaweed bed images from a drone perspective, the method of this paper significantly improves segmentation accuracy and enhances the model's generalization capabilities, providing strong technical support for the monitoring and protection of seaweed beds. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the present invention or the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, what is described below is only one embodiment of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0035] Figure 1 This is the overall structural block diagram of the image segmentation system of the present invention.

[0036] Figure 2 This is the network structure diagram of the improved ResNet-50 in the present invention.

[0037] Figure 3 It is a structural diagram of the group mapping module GEPGMM of the present invention.

[0038] Figure 4 This is the network structure diagram of the channel space feature fusion module CSFFM of the present invention.

[0039] Figure 5 This is a schematic diagram of the segmentation results of the present invention on the UAV seaweed bed image. DETAILED DESCRIPTION

[0040] The invention will be further described below with reference to specific embodiments.

[0041] The present invention proposes an image segmentation system and method for UAV seaweed bed monitoring. The system structure diagram is shown in FIG. Figure 1 As shown in the figure, based on the Iterative Few-shot Adaptor framework, a segmentation network model that combines channel-space feature fusion is constructed. It can be divided into three parts: feature grouping and mapping, feature fusion, and segmentation prediction. UAV-captured seaweed bed images are used as input; the seaweed bed image data captured by the drone is divided into a support image set and a query image set. The support image set contains annotated images, while the query image set contains only unannotated images. The data processing process includes:

[0042] Feature grouping and mapping part: First, the improved ResNet-50 model is used to extract features from the input image, and then the extracted features are grouped and fused using the Gaussian Energy Prototype-driven Grouping and Mapping Module (GEPGMM). ResNet-50, with its powerful feature extraction capabilities and residual structure design, can effectively process complex information in images while avoiding the gradient vanishing problem in deep network training. In the present invention, the ResNet-50 model is appropriately modified to remove the fully connected layer and the last residual block (Stage5) in the original network, while retaining the first four stages (Stage1-Stage4) to extract multi-scale feature maps. These feature maps not only contain shallow detail information of the image, but also cover deep semantic information, providing a rich feature foundation for subsequent feature fusion and segmentation tasks.

[0043] Feature Fusion: The features generated by the feature grouping and mapping phase are grouped and fused using the Channel Spatial Feature Fusion Module (CSFFM). The CSFFM further fuses the features generated by the GEPGMM using channel and spatial attention mechanisms, enhancing feature representation. The channel attention mechanism enables the model to focus on the most important channel features, while the spatial attention mechanism focuses on key spatial locations in the image. This allows the model to more accurately identify and focus on key feature regions in the image, thereby improving segmentation accuracy and robustness.

[0044] Segmentation prediction part: Finally, the segmentation prediction part includes a decoder, which uses three segmentation heads to convert the fused feature maps into the final segmentation results. The segmentation head generates a segmentation mask with the same resolution as the input image through convolution operations and activation functions. During training, the output of the segmentation head is compared with the true mask label, and the difference between the predicted result and the true label is calculated using the cross-entropy loss function, thereby guiding the model to optimize the segmentation result. At the same time, a cross-domain matching loss constraint is introduced to support the alignment of domain and query domain features, ensuring that the model has good generalization ability in cross-domain tasks. During training, the model is first pre-trained on the source domain dataset to learn general feature representation and segmentation capabilities. Then, it is fine-tuned on the target domain dataset to adapt to the specific feature distribution of the target domain. Through this two-stage training strategy, the model can achieve effective knowledge transfer between the source and target domains, significantly improving the accuracy and robustness of cross-domain small-sample segmentation.

[0045] 1. About the Improved ResNet-50 Model

[0046] In this invention, the feature grouping mapping part uses the encoder, namely the ResNet-50 model, for feature extraction because of its strong feature extraction capability and good generalization performance. The ResNet-50 model is structurally modified to remove the fully connected layer and the last residual block (Stage 5), while retaining the first four stages (Stage 1-Stage 4). Through this modification, the model can extract multi-scale feature maps while reducing the amount of computation and the number of parameters. The structure is as follows Figure 2 As shown in the figure, the feature map output by Stage 1 has a resolution of approximately 1 / 4 that of the input image and contains rich details. The feature map output by Stage 2 has a resolution of approximately 1 / 8 that of the input image and contains medium-scale features. The feature map output by Stage 3 has a resolution of approximately 1 / 16 that of the input image and contains deeper semantic information. The feature map output by Stage 4 has a resolution of approximately 1 / 32 that of the input image and contains the deepest semantic information. Next, the feature extraction process begins. The normalized input image (size 512×512) is fed into the ResNet-50 model. The input image first passes through the first convolutional layer and the max pooling layer, halving the image size and extracting shallow, detailed features to produce the feature map for Stage 1. The second residual block then further extracts mid-level features and halves the resolution again to produce the feature map for Stage 2. The third residual block then extracts deeper semantic features and halves the resolution again to produce the feature map for Stage 3. Finally, the fourth residual block extracts the deepest semantic features and further halves the resolution to produce the feature map for Stage 4. Finally, the feature grouping and mapping component outputs feature maps. These feature maps encompass a rich set of information, from shallow details to deep semantics, providing a solid foundation for subsequent feature fusion and segmentation tasks. To prepare for feature fusion, appropriate feature maps (such as P3 and P4) are selected from Stages 1-4 as input to the Gaussian energy prototype-driven grouping and mapping module. These feature maps not only contain rich details but also deep semantic information, providing high-quality feature representation for subsequent feature fusion and segmentation tasks.

[0047] 2. About the Grouping and Mapping Module GEPGMM

[0048] The Gaussian Energy Prototype-Driven Grouping and Mapping Module (GEPGMM) is a key module in the present invention. Its structure is as follows: Figure 3 As shown in the figure, the goal is to use Gaussian energy to measure similarity, group and map similar features, and generate more representative prototypes, thereby more effectively capturing and optimizing foreground object features. The specific steps are as follows:

[0049] First, use the Canny edge detection algorithm to extract edge features from the support image and generate an edge map to mark the locations of edge points. From these edge points, representative seed points are randomly sampled, for example, 50 edge points are randomly selected as initial seed points. These seed points serve as the starting point for subsequent feature grouping, ensuring diversity and representativeness in the grouping process.

[0050] Next, for each selected seed point, the corresponding feature vector is extracted from the feature map Fs of the support image. These feature vectors contain rich semantic information and provide a foundation for subsequent feature grouping and prototype generation. The feature vectors are obtained from feature maps extracted from a stage of the ResNet-50 model (such as Stage 4), which have high semantic expressiveness.

[0051] Next, the feature distance Dfc and spatial distance Dsc between each pixel in the support image feature map Fs and the seed point are calculated. Feature distance Dfc can be measured using cosine similarity, while spatial distance Dsc can be calculated using Euclidean distance. The smoothing term and energy function are combined to update the feature point center. The specific formula is as follows:

[0052]

[0053] in, Dfc ( p , Ci )=1−cos( θp , Ci ) 、θp , Ci is the eigenvector Fs ( p )and Ci The angle between Dsc ( p , Ci )=∥ p − Ci ∥2. λ is the weight coefficient, which is used to balance the influence of feature distance and spatial distance.

[0054] Through an iterative optimization process (e.g., 5 iterations), the feature point centers are gradually updated, and finally 5 sets of multi-scale prototypes are generated. These prototypes can cover target features of different sizes and provide richer feature representations for subsequent segmentation tasks.

[0055] The generated prototype is then expanded and mapped to the query image for segmentation. The most appropriate prototype is dynamically selected by calculating the similarity between the expanded feature points and the query features. The similarity calculation formula is as follows:

[0056]

[0057] in, is the feature of pixel p in the query image, is the center of the i-th feature point, is the standard deviation of the Gaussian function, which is used to control the decay rate of the similarity.

[0058] Finally, a response map (respmap) and a guide map (guidemap) are generated. The response map (respmap) represents the maximum similarity between each pixel in the query image and the prototype, while the guide map (guidemap) represents the prototype index corresponding to each pixel. These two maps are concatenated with the original query feature map (Fq) to generate the optimized query feature map (a2). This process provides richer feature information for subsequent segmentation tasks and enhances the model's ability to recognize foreground objects.

[0059] Through the above steps, the GEPGMM module can effectively group and map similar features, generating more representative prototypes that better capture and optimize the characteristics of foreground objects. This not only improves the model's segmentation accuracy in cross-domain small-sample segmentation tasks, but also enhances its generalization ability, enabling it to maintain high segmentation performance across different domains and conditions.

[0060] 3. About the Channel Space Feature Fusion Module CSFFM

[0061] The channel-space feature fusion module CSFFM (Channel-Space Feature Fusion Module) is a key module in this invention, which aims to fuse feature maps through channel and spatial attention mechanisms, thereby enhancing feature representation capabilities and improving segmentation accuracy. Its structure is as follows Figure 4 Specifically, the CSFFM module starts from the feature maps output by the Gaussian Energy Prototype Driven Group Mapping Module (GEPGMM). Although these feature maps contain rich semantic and spatial information, they need to be further optimized to improve the segmentation performance.

[0062] First, the module uses a channel-wise attention mechanism to enhance the features of important channels in the feature map while suppressing unimportant channels. This process is achieved through global average pooling (GAP) and global max pooling (GMP), respectively, resulting in two feature vectors A and M, with shapes (C, 1, 1), where C is the number of channels in the feature map. These two vectors are then concatenated and passed to a 1×1 convolutional layer to generate channel-wise attention weights Wc. The sigmoid activation function σ is used to normalize the weights to the range (0, 1) to emphasize important channels and suppress unimportant ones. The specific formula is as follows:

[0063]

[0064]

[0065]

[0066] Next, the module uses a spatial attention mechanism to enhance the features of important spatial locations in the feature map while suppressing unimportant locations. This process first calculates the mean μ and maximum max of the input feature map FGEPGMM in the channel dimension, which has a shape of (1, H, W), where H and W are the height and width of the feature map, respectively. Then, μ and max are concatenated and passed to a 3×3 convolutional layer to generate the spatial attention weights Ws, which are also normalized using the sigmoid activation function σ. The specific formula is as follows:

[0067]

[0068]

[0069] Finally, the channel attention weight Wc and the spatial attention weight Ws are added to obtain the final combined attention weight, and the input feature map FGEPGMM is weighted to generate the optimized feature map F′. This process is achieved through element-by-element multiplication ⊗, and the specific formula is as follows:

[0070]

[0071] Through this dual-dimensional feature fusion, the CSFFM module not only improves the model's ability to identify foreground objects but also significantly enhances segmentation accuracy. The channel attention mechanism emphasizes the features of important channels while suppressing less important ones, while the spatial attention mechanism emphasizes the features of important spatial locations while suppressing less important ones. This fusion approach provides richer feature information for subsequent segmentation tasks, enhancing the model's overall performance.

[0072] IV. Experimental Results

[0073] In the experiment of seaweed bed image segmentation task, the network model proposed in this paper showed excellent performance, and the visualization results are as follows: Figure 5As shown in Table 1. On the seaweed bed dataset, the model achieved significant performance improvements. The mean intersection over union (mIoU) in the 1-shot and 5-shot scenarios reached 72.5% and 76.8%, respectively, representing improvements of 3.0% and 3.8% compared to the baseline methods. This result not only demonstrates the efficiency and accuracy of CDGCNet in handling seaweed bed image segmentation tasks, but also highlights its strong generalization capabilities in small-sample segmentation scenarios. By introducing the Gaussian Energy Prototype-Driven Group Mapping Module (GEPGMM) and the Channel Spatial Feature Fusion Module (CSFFM), the model can effectively capture and optimize the key features of the seaweed bed area while reducing background interference. These innovative designs give CDGCNet a significant advantage in the field of seaweed bed image segmentation, providing strong technical support for marine ecological monitoring and protection.

[0074] Table 1 Segmentation results on self-made dataset

[0075]

[0076] The above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

[0077] Although the above describes the specific implementation methods of the present invention, it does not limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. An image segmentation system for UAV seaweed bed monitoring, characterized by: The improved method is based on the iterative small sample adapter (IFA), which takes the seaweed bed image taken by the UAV as input, including the feature grouping mapping part, the feature fusion part and the segmentation prediction part. The seaweed bed image data taken by drones is divided into a support image set and a query image set. The support image set contains annotated images, while the query image set contains only unannotated images. The feature group mapping part includes an encoder and a Gaussian energy prototype-driven group mapping module (GEPGMM), which is used to extract features and strengthen the focus on key features. The encoder uses the ResNet-50 model, removes the fully connected layer and the last residual block of the original network, and uses the last three network layers of different depths to extract feature information from the input image to obtain the original query feature map Fq and the support image feature map Fs. The group mapping module GEPGMM includes a seed point selection unit and a feature vector extraction unit. The seed point selection unit first uses the Canny edge detection algorithm to extract the edge features of the support image set image, and then selects representative seed points by random sampling; the feature vector extraction unit extracts the feature vector of the corresponding support image feature map Fs for each seed point, and then calculates the feature distance and spatial distance, and updates the feature point center in combination with the Gaussian function to finally generate a multi-scale prototype; at the same time, a response map respmap and a guide map guidemap are generated, and spliced ​​with the original query feature map Fq to generate an optimized query feature map a2; The feature fusion part adopts the channel space feature fusion module CSFFM to perform channel and spatial dimension feature fusion on the query feature map a2 extracted by the feature grouping mapping part. The CSFFM module includes a channel attention unit and a spatial attention unit, wherein the channel attention unit generates channel weights through global pooling operation and convolution, and the spatial attention unit generates spatial weights through convolution operation; Add the channel weight and spatial weight to weight the input feature map to obtain the optimized feature map; The segmentation prediction part includes a decoder, which uses three segmentation heads and is responsible for converting the optimized feature map into the final segmentation mask.

2. The image segmentation system for monitoring seaweed beds using an unmanned aerial vehicle (UAV) as claimed in claim 1, characterized in that: The channel space feature fusion module CSFFM, for the input feature map, where the number of channels, height and width of the feature map are C, H and W respectively, the channel attention module obtains the aggregated features through global average pooling and global maximum pooling, and then uses shared 1D convolution to generate channel weights ; The spatial attention module generates spatial weights through 3×3 convolution ,Finally, the channel weight and spatial weight are added together to weight the input feature map to obtain the optimized feature map.

3. The image segmentation system for seaweed bed monitoring using an unmanned aerial vehicle according to claim 1, wherein: The encoder adopts the ResNet-50 model, removes the fully connected layer and the last residual block Stage5, and retains the first four stages Stage1-Stage4; the resolution of the feature map output by Stage1 is about 1 / 4 of the input image, containing rich detail information; the resolution of the feature map output by Stage2 is about 1 / 8 of the input image, containing medium-scale features; the resolution of the feature map output by Stage3 is about 1 / 16 of the input image, containing deeper semantic information; the resolution of the feature map output by Stage4 is about 1 / 32 of the input image, containing the deepest semantic information; then, the feature extraction process is carried out, and the normalized input image is sent to ResNet-5 0 model, the input image first passes through the first convolutional layer and the maximum pooling layer, the image size is halved, and the shallow detail features are extracted to obtain the feature map of Stage1. Then, the middle-level features are further extracted through the second residual block, and the resolution is halved again to obtain the feature map of Stage2. Then, the deeper semantic features are extracted through the third residual block, and the resolution is halved again to obtain the feature map of Stage3. Finally, the deepest semantic features are extracted through the fourth residual block, and the resolution is further halved to obtain the feature map of Stage4. Finally, the feature grouping mapping part outputs the feature map; the appropriate feature map is selected from Stage1-Stage4 as the input of the grouping mapping module GEPGMM.

4. The image segmentation system for monitoring seaweed beds using an unmanned aerial vehicle according to claim 1, wherein: The specific data processing process of the group mapping module GEPGMM is as follows: S1, use the Canny edge detection algorithm to extract the edge features of the support image set, generate an edge map to mark the location of the edge points, and randomly sample representative seed points from the edge points as the starting point for subsequent feature grouping; S2, for each selected seed point, extract the corresponding feature vector from the support image feature map Fs. The feature vector is obtained from the feature map extracted from a certain stage of the ResNet-50 model; S3, calculate the feature distance Dfc and spatial distance Dsc between each pixel in the support image feature map Fs and the seed point; the feature distance Dfc is measured by cosine similarity, while the spatial distance Dsc is calculated by Euclidean distance, combined with the smoothing term and Gaussian function, to update the feature point center; S4, through the iterative optimization process, gradually updates the feature point centers, and finally generates 5 sets of multi-scale prototypes, which can cover target features of different sizes; S5, expands and maps the generated prototypes into the query image for segmentation, and dynamically selects the most appropriate prototype by calculating the similarity between the expanded feature points and the query features; S6, generate the response map respmap and the guide map guidemap; the response map respmap represents the maximum similarity between each pixel in the query image and the prototype, while the guide map guidemap represents the prototype index corresponding to each pixel. These two maps are spliced ​​with the original query feature map Fq to generate the optimized query feature map a2.

5. The image segmentation system for monitoring seaweed beds using an unmanned aerial vehicle according to claim 2, wherein: The specific data processing process of the channel space feature fusion module CSFFM is as follows: First, the channel attention mechanism is used to enhance the features of important channels in the query feature map a2, while suppressing unimportant channels. This is achieved through global average pooling (GAP) and global maximum pooling (GMP), respectively, to obtain two feature vectors A and M, whose shapes are (C, 1, 1), where C is the number of channels in the feature map. These two vectors are then concatenated and passed to a 1×1 convolutional layer to generate channel attention weights Wc. The Sigmoid activation function σ is used to normalize the weights to the range of (0, 1) to highlight important channels and suppress unimportant channels. Next, the spatial attention mechanism is used to enhance the features of important spatial locations in the query feature map a2 output by the group mapping module, while suppressing unimportant locations. First, the mean μ and maximum max of the input query feature map a2 in the channel dimension are calculated, and its shape is (1, H, W), where H and W are the height and width of the feature map respectively. Then, μ and max are concatenated and passed to a 3×3 convolutional layer to generate the spatial attention weight Ws, which is also normalized using the Sigmoid activation function σ. Finally, the channel attention weight Wc and the spatial attention weight Ws are added to obtain the final combined attention weight, and the query feature map a2 is weighted to generate the optimized feature map F′.

6. An image segmentation method for UAV seaweed bed monitoring, characterized in that: The following processes are included: Obtain image data of seaweed beds in target areas through drone photography; Inputting image data into the image segmentation system according to any one of claims 1 to 5 for segmenting a target area of ​​the input image; By performing real-time online analysis on image data and outputting segmentation results, the target area can be accurately segmented.

Images