Turtle instance segmentation method and system based on weight low-rank decomposition and feature enhancement

Abstract

The application belongs to the technical field of image processing, and particularly relates to a turtle instance segmentation method and system based on weight low-rank decomposition and feature enhancement, which comprises the following steps: acquiring an underwater turtle image; constructing an image encoder based on a weight low-rank decomposition and updating mechanism, capturing global features of the acquired turtle image according to the image encoder and a multi-head self-attention mechanism; performing selective feature enhancement on the obtained global features of the turtle image to obtain a multi-scale reconstruction feature map; generating a candidate region on the obtained multi-scale reconstruction feature map, aligning the multi-scale reconstruction feature map on the generated candidate region, detecting and segmenting the aligned multi-scale reconstruction feature map through a multi-task branch network, and completing turtle instance segmentation.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a turtle instance segmentation method and system based on weighted low-rank decomposition and feature enhancement. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Sea turtles are an endangered species. Sea turtle instance segmentation can be used to investigate sea turtle diversity and conduct statistical analysis, further facilitating sea turtle conservation research. However, sea turtles live in complex underwater environments where the water medium and suspended particles significantly absorb and reflect light, resulting in low contrast and blurred edges in the acquired underwater sea turtle images. This severely affects the feature representation of sea turtle images.

[0004] Currently, traditional network models cannot selectively enhance weak turtle features when segmenting underwater turtle targets, which greatly reduces the accuracy of turtle instance segmentation. At the same time, directly updating large weight matrices during model training results in low training efficiency and weak model adaptability, making it unable to quickly and accurately segment turtle instances in complex underwater environments. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a turtle instance segmentation method and system based on weighted low-rank decomposition and feature enhancement. By designing a weighted low-rank decomposition and update mechanism, the adaptability and training efficiency of the network are improved. Furthermore, an innovative turtle feature selective enhancement module is designed to effectively enhance the network's ability to perceive features in blurred images and improve the segmentation accuracy of underwater turtles.

[0006] According to some embodiments, the first solution of the present invention provides a turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement, employing the following technical solution:

[0007] A turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement includes:

[0008] Acquire underwater images of sea turtles;

[0009] An image encoder is constructed based on a weighted low-rank decomposition and update mechanism, and global features of the acquired turtle images are captured based on the image encoder and a multi-head self-attention mechanism.

[0010] Feature-selective enhancement is performed on the global features of the obtained sea turtle images to obtain multi-scale reconstructed feature maps;

[0011] Candidate regions are generated on the obtained multi-scale reconstructed feature maps. The multi-scale reconstructed feature maps are aligned on the generated candidate regions. The aligned multi-scale reconstructed feature maps are then detected and segmented through a multi-task branch network to complete the turtle instance segmentation.

[0012] As a further technical limitation, the constructed image encoder is equipped with a weighted low-rank decomposition and update module; in the weighted low-rank decomposition and update module, the weights are decomposed into low-rank matrices through a weighted low-rank decomposition and update mechanism, and the update is completed.

[0013] As a further technical limitation, in the feature selective enhancement, a turtle image feature map is obtained based on the global features and selectivity coefficients of the turtle image. The obtained turtle image feature map is iterated layer by layer to obtain a multi-level feature map. The obtained multi-level feature maps are fused to obtain a fused feature map. Convolution and pooling are used to reconstruct the obtained fused feature map at different scales to obtain a multi-scale reconstructed feature map.

[0014] As a further technical limitation, the turtle instance segmentation includes a candidate region generation stage, an alignment stage, and a multi-task prediction stage; wherein, in the candidate region generation stage, the region generation network introduces anchor boxes of a preset size on feature points, and based on the confidence of the target's existence predicted by the anchor boxes and the regression offset of the bounding boxes, non-maximum suppression is used to filter all candidate boxes to obtain candidate regions, thus completing the candidate region generation stage.

[0015] Furthermore, in the alignment stage, bilinear interpolation is used to obtain feature values ​​at the sampling positions, and max pooling is combined to extract fixed-size region features; the aligned features are then fed into a multi-task branch network to complete the detection and segmentation tasks of the scale-reconstructed feature map.

[0016] As a further technical limitation, the loss function for turtle instance segmentation includes a region proposal network loss function, a detector classification loss function, a detector bounding box regression loss function, and a segmentation mask loss function.

[0017] According to some embodiments, the second aspect of the present invention provides a turtle instance segmentation system based on weighted low-rank decomposition and feature enhancement, employing the following technical solution:

[0018] A turtle instance segmentation system based on weighted low-rank decomposition and feature enhancement includes:

[0019] The image acquisition module is fully configured to acquire underwater sea turtle images;

[0020] The feature capture module is fully configured to construct an image encoder based on a weighted low-rank decomposition and update mechanism, and capture the global features of the acquired turtle images based on the image encoder and the multi-head self-attention mechanism.

[0021] The feature enhancement module is configured to selectively enhance the global features of the obtained turtle image to obtain a multi-scale reconstructed feature map.

[0022] The instance segmentation module is configured to generate candidate regions on the obtained multi-scale reconstructed feature maps, align the multi-scale reconstructed feature maps on the generated candidate regions, and perform detection and segmentation of the aligned multi-scale reconstructed feature maps through a multi-task branch network to complete turtle instance segmentation.

[0023] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium, employing the following technical solution:

[0024] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement as described in the first aspect of the present invention.

[0025] According to some embodiments, the fourth aspect of the present invention provides an electronic device, which adopts the following technical solution:

[0026] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps in the turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement as described in the first aspect of the present invention.

[0027] According to some embodiments, the fifth aspect of the present invention provides a computer program product, which adopts the following technical solution:

[0028] A computer program product includes software code, wherein the program in the software code performs the steps of the turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement as described in the first aspect of the present invention.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] This invention improves the adaptability and training efficiency of the network by designing a weighted low-rank decomposition and update mechanism, and innovatively designs a turtle feature selective enhancement module to effectively enhance the network's ability to perceive features in blurred images and improve the segmentation accuracy of underwater turtles. Attached Figure Description

[0031] The accompanying drawings, which form part of this embodiment, are used to provide a further understanding of this embodiment. The illustrative embodiments and their descriptions are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0032] Figure 1 This is a flowchart of the turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement in Embodiment 1 of the present invention;

[0033] Figure 2 This is a network architecture diagram of the turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement in Embodiment 1 of the present invention.

[0034] Figure 3 This is a schematic diagram of the image encoder based on weighted low-rank decomposition update in Embodiment 1 of the present invention;

[0035] Figure 4 This is a schematic diagram of the weighted feature fusion structure in Embodiment 1 of the present invention;

[0036] Figure 5 This is a schematic diagram of the turtle instance segmentation head in Embodiment 1 of the present invention;

[0037] Figure 6 This is a schematic diagram of the turtle instance segmentation result in Embodiment 1 of the present invention;

[0038] Figure 7 This is a structural block diagram of the turtle instance segmentation system based on weighted low-rank decomposition and feature enhancement in Embodiment 2 of the present invention. Detailed Implementation

[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0041] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0042] In this invention, terms such as "upper," "lower," "left," "right," "front," "back," "vertical," "horizontal," "side," and "bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used only to facilitate the description of the structural relationships of the various components or elements of this invention and do not specifically refer to any component or element in this invention. They should not be construed as limiting the invention.

[0043] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0044] Example 1

[0045] Embodiment 1 of this invention introduces a turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement.

[0046] like Figure 1 The turtle instance segmentation method shown includes:

[0047] Acquire underwater images of sea turtles;

[0048] An image encoder is constructed based on a weighted low-rank decomposition and update mechanism, and global features of the acquired turtle images are captured based on the image encoder and a multi-head self-attention mechanism.

[0049] Feature-selective enhancement is performed on the global features of the obtained sea turtle images to obtain multi-scale reconstructed feature maps;

[0050] Candidate regions are generated on the obtained multi-scale reconstructed feature maps. The multi-scale reconstructed feature maps are aligned on the generated candidate regions. The aligned multi-scale reconstructed feature maps are then detected and segmented through a multi-task branch network to complete the turtle instance segmentation.

[0051] like Figure 2 As shown, the network in this embodiment includes: an image encoder based on weighted low-rank decomposition and updating, which improves the model's adaptability and segmentation ability in underwater images by designing a weighted low-rank decomposition and updating mechanism, while reducing the computational resources and time required for training; a turtle feature selective enhancement module, which consists of a key feature selection module and a multi-scale feature generator. The feature selection module selects key features output by the image encoder and performs multi-level weighted fusion; the multi-scale feature generator performs multi-scale splitting and generation on the fused feature map; and a turtle instance segmentation head, which uses a built-in region generation network to generate candidate regions, and after passing through a region alignment module, the actual candidate regions are accurately aligned with the original image, achieving accurate instance segmentation.

[0052] As one or more implementations, image encoders based on weighted low-rank decomposition updates utilize multi-head self-attention mechanisms to capture global contextual information in images, thereby extracting high-level feature representations. For example... Figure 3 As shown, the encoder mainly consists of five modules: an image patch encoding module, a position encoding module, an encoder module, a feature fusion module, and a weighted low-rank decomposition update module. The image patch encoding module divides the image into 16×16 image patches and uses a 2D convolution operation with a kernel size and stride of 16 to map each image patch to a 768-dimensional feature vector; that is...

[0053] ;

[0054] in, I Indicates the input image. This represents the convolution operation. N This represents the total number of image blocks.

[0055] The location encoding module adds a learnable location vector to each image patch to preserve spatial location information, enabling the encoder to model the structural features of the image. The encoder module utilizes a multi-head self-attention mechanism and a feedforward neural network to model the complex relationships between global and local features. The multi-layer encoder can extract high-dimensional semantic features at multiple levels, providing rich and diverse feature representations for subsequent segmentation tasks; that is...

[0056] ;

[0057] ;

[0058] ;

[0059] in, These represent the query value, key value, and raw value matrices, respectively, all derived from the input features. It is the first i The linear projection matrix of each attention head. The output projection matrix used to fuse all attention head outputs; It is the feature dimension of each attention head, usually 1. ,in h It's about the number of heads; This is the attention weight, used to measure the similarity between the query value and the key value. The feature fusion module fuses and optimizes the features extracted by the encoder.

[0060] Because the encoder module employs a multi-head self-attention mechanism, it suffers from high computational complexity. To improve the network's segmentation performance and reduce computational complexity, this embodiment designs a low-rank weight decomposition and update mechanism for lightweight and efficient fine-tuning of the encoder. The core idea of ​​this mechanism is to introduce a low-rank parameterization method, performing low-rank decomposition on some weights of the pre-trained model and fine-tuning the resulting low-rank matrix to adapt to specific downstream tasks; such as... Figure 3 As shown, input For the pre-trained weight matrix and The update is constrained by using low-rank decomposition; that is...

[0061] ;

[0062] ;

[0063] in, , , and rank During training, the weight matrix is ​​frozen. and Only for A and B The parameters in the algorithm are updated using gradients.

[0064] This embodiment selects uniform distribution initialization. A The parameters, and B All parameters are set to zero. Compared to directly adjusting all weight parameters, fine-tuning greatly reduces computational costs, and good fine-tuning results can be obtained by adjusting only a few parameters.

[0065] Due to issues such as color shift, low illumination, and image blurring in underwater images, image encoders struggle to accurately capture the edge details and complete semantic information of sea turtles, thus affecting image segmentation accuracy. To address this, this implementation presents a sea turtle feature-selective enhancement module for underwater blurred image segmentation. This module consists of two core parts: a key feature selection module and a multi-scale feature generator. For example... Figure 4 As shown, the multi-stage selection mechanism of the key feature selection module can achieve multi-level feature interaction fusion, which can significantly enhance the semantic information of underwater images; its selective enhancement process is as follows:

[0066] ;

[0067] ;

[0068] ;

[0069] ;

[0070] ;

[0071] in, This represents the features generated by the encoder. express go through Features after downsampling; application of upper-layer features Generate selectivity coefficient Used to highlight key features; selectivity coefficient With low-level feature maps Perform element-wise multiplication and then combine with the high-level feature map. Add them together to get the updated feature map. After layer-by-layer iteration, the selectively enhanced multi-layer feature maps are fused into semantically rich feature information, and then... Restore the channel dimension to generate the final fused feature map. ;right Convolution and pooling operations are used to reconstruct feature maps at different scales, generating five feature maps of different scales for subsequent instance segmentation tasks.

[0072] The turtle instance segmentation head structure is as follows: Figure 5 As shown, the process includes three stages: candidate region generation, precise alignment, and multi-task prediction. First, the region generation network generates candidate regions on multi-scale feature maps. Specifically, the region generation network introduces three anchor boxes of preset sizes and aspect ratios at each feature point, and simultaneously predicts the confidence of the target's presence and the regression offset of the bounding box for each anchor box. Non-maximum suppression is used to filter all candidate boxes, ultimately retaining 1000 high-quality candidate regions as input for subsequent processing. The region alignment module precisely aligns the candidate regions with the feature maps, avoiding quantization errors caused by traditional alignment methods. This module uses bilinear interpolation to obtain feature values ​​of four sub-pixel points at each sampling location, and combines max pooling to extract fixed-size region features. The aligned features are then fed into a multi-task branch network to complete detection and segmentation tasks respectively. One branch is responsible for target classification and bounding box regression, used to further identify target categories and refine their spatial locations; the other branch uses a fully convolutional network structure to generate a high-resolution binary mask corresponding to the target, achieving pixel-level instance segmentation. This design not only ensures the accuracy of detection and segmentation, but also has strong generalization ability, making it suitable for target segmentation tasks in complex backgrounds in underwater images.

[0073] As one or more implementation methods, after the network structure in this embodiment is built, a loss function needs to be designed to train the network. The loss function of the underwater turtle instance segmentation network consists of region proposal network loss, detector classification loss, detector bounding box regression loss, and segmentation mask loss; the total loss function... for: ;

[0074] The candidate region generation network is responsible for generating candidate regions. Its loss includes classification loss (foreground / background) and bounding box regression loss. The former uses cross-entropy loss, and the latter uses smooth L1 loss. The loss function is as follows:

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] in, These are anchor tags, representing the background and foreground respectively; It is the probability of predicting it as the prospect. It is a normalization term; For the predicted bounding box parameters, These are the parameters of the actual bounding box; This is the balance coefficient; The number of positive samples serves as a normalization factor.

[0080] The classification loss of the detection head is used to predict the target category of the candidate region, and the multi-class cross-entropy loss function is employed. for ;in, This serves as a normalization function for the number of candidate regions. The true category label for the candidate region. This is the probability of predicting the true category.

[0081] The bounding box regression loss of the detection head is used for fine-tuning the bounding boxes of positive samples. A smooth L1 loss is employed to measure the difference between the predicted bounding box parameters and the true bounding boxes, optimizing the model to generate more accurate target localization results. ;in, and These are the bounding box parameters for the predicted and ground truth boxes, respectively. This represents the number of positive samples.

[0082] The segmentation mask loss uses pixel-wise binary cross-entropy to measure the difference between the predicted mask and the true mask. For each positive sample, only the binary mask loss of its class is calculated, thereby optimizing the model to generate more accurate segmentation results; its segmentation mask loss function... for: ;in, For the true mask at position ( h , w The tag; The predicted mask probability; This represents the total number of pixels in the positive samples.

[0083] During the training phase, data augmentation was performed using random horizontal flipping, random scaling, and random cropping. The experimental environment consisted of Python 3.10 (Ubuntu 20.04), PyTorch 2.1.1, and CUDA 12.1. All experiments were conducted using a single NVIDIA GeForce RTX 4090 PLUS GPU. Training parameters were set as follows: optimizer: AdamW, initial learning rate: 0.0001, weight decay rate: 0.1, batch size: 4, training epochs: 24; image resolution: 1024×1024.

[0084] In this embodiment, the proposed segmentation network was trained using a specially collected dataset of underwater sea turtle images, and its performance was evaluated on the test set of the same dataset; the experimental results are as follows. Figure 6 As shown, the proposed network performs excellently in both object detection and instance segmentation tasks. On the test set of the Turtle dataset, the average precision (AP) of the proposed network for bounding boxes is [missing information]. box The mask average accuracy reached 88.8%. mask The accuracy rate reached 87.7%, verifying that the network has good target detection and segmentation capabilities in complex underwater environments.

[0085] This embodiment improves the network's adaptability and training efficiency by designing a weighted low-rank decomposition and update mechanism, and innovatively designs a turtle feature selective enhancement module to effectively enhance the network's ability to perceive blurred image features and improve the segmentation accuracy of underwater turtles.

[0086] Example 2

[0087] Embodiment 2 of this invention introduces a turtle instance segmentation system based on weighted low-rank decomposition and feature enhancement.

[0088] like Figure 7 The turtle instance segmentation system shown includes:

[0089] The image acquisition module is fully configured to acquire underwater sea turtle images;

[0090] The feature capture module is fully configured to construct an image encoder based on a weighted low-rank decomposition and update mechanism, and capture the global features of the acquired turtle images based on the image encoder and the multi-head self-attention mechanism.

[0091] The feature enhancement module is configured to selectively enhance the global features of the obtained turtle image to obtain a multi-scale reconstructed feature map.

[0092] The instance segmentation module is configured to generate candidate regions on the obtained multi-scale reconstructed feature maps, align the multi-scale reconstructed feature maps on the generated candidate regions, and perform detection and segmentation of the aligned multi-scale reconstructed feature maps through a multi-task branch network to complete turtle instance segmentation.

[0093] The detailed steps are the same as those of the turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement provided in Example 1, and will not be repeated here.

[0094] Example 3

[0095] Embodiment 3 of the present invention provides a computer-readable storage medium.

[0096] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement as described in Embodiment 1 of the present invention.

[0097] The detailed steps are the same as those of the turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement provided in Example 1, and will not be repeated here.

[0098] Example 4

[0099] Embodiment 4 of the present invention provides an electronic device.

[0100] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements the steps in the turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement as described in Embodiment 1 of the present invention.

[0101] The detailed steps are the same as those of the turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement provided in Example 1, and will not be repeated here.

[0102] Example 5

[0103] Embodiment 5 of the present invention provides a computer program product.

[0104] A computer program product includes software code, wherein the program in the software code performs the steps of the turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement as described in Embodiment 1 of the present invention.

[0105] The detailed steps are the same as those of the turtle instance segmentation method based on weighted low-rank decomposition and feature enhancement provided in Example 1, and will not be repeated here.

[0106] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0107] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0108] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0109] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0110] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0111] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0112] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A turtle instance segmentation method based on weight low-rank decomposition and feature enhancement, characterized in that, The method comprises the following steps: obtaining an underwater turtle image; constructing an image encoder based on a weight low-rank decomposition and updating mechanism, specifically, designing a weight low-rank decomposition and updating mechanism, performing lightweight and efficient fine-tuning on the encoder, introducing a low-rank parameterization method, decomposing part of the weights of the pre-trained model into low-rank matrices, and fine-tuning the decomposed low-rank matrices to adapt to specific downstream tasks; capturing global features of the obtained turtle image according to the image encoder and a multi-head self-attention mechanism; performing feature selective reinforcement on the obtained global features of the turtle image to obtain a multi-scale reconstructed feature map; generating a candidate region on the obtained multi-scale reconstructed feature map, aligning the multi-scale reconstructed feature map on the generated candidate region, and performing detection and segmentation of the multi-scale reconstructed feature map after alignment through a multi-task branch network to complete turtle instance segmentation.

2. The method of claim 1, wherein the method is based on weight low-rank decomposition and feature enhancement. The constructed image encoder is internally provided with a weight low-rank decomposition and updating module; in the weight low-rank decomposition and updating module, the weights are decomposed into low-rank matrices through the weight low-rank decomposition and updating mechanism, and the updating is completed.

3. The method of claim 1, wherein the method is based on weight low-rank decomposition and feature enhancement. In the feature selective reinforcement, based on the global features of the turtle image and a selective coefficient, a turtle image feature map is obtained, the obtained turtle image feature map is iterated layer by layer to obtain a multi-level feature map, the obtained multi-level feature map is fused to obtain a fused feature map, and the fused feature map is reconstructed at different scales through convolution and pooling to obtain a multi-scale reconstructed feature map.

4. The method of claim 1, wherein the method is based on weight low-rank decomposition and feature enhancement. The turtle instance segmentation comprises a candidate region generation stage, an alignment stage and a multi-task prediction stage; in the candidate region generation stage, an anchor box of a preset size is introduced on a feature point by a region generation network, a confidence degree of the existence of a target and a regression offset of a bounding box are predicted based on the anchor box, all candidate boxes are screened through non-maximum suppression to obtain a candidate region, and the generation stage of the candidate region is completed.

5. The method of claim 4, wherein the method is based on weight low-rank decomposition and feature enhancement. In the alignment stage, feature values are obtained on a sampling position through bilinear interpolation, and region features of a fixed size are extracted through a max-pooling operation; the aligned features are sent into a multi-task branch network to respectively complete the detection and segmentation tasks of the scale reconstructed feature map.

6. The method of claim 1, wherein the method is based on weight low-rank decomposition and feature enhancement. The loss function of the turtle instance segmentation comprises a region proposal network loss function, a detection head classification loss function, a detection head bounding box regression loss function and a segmentation mask loss function.

7. A turtle instance segmentation system based on weight low-rank decomposition and feature enhancement, characterized in that, The method comprises the following steps: an image acquisition module is configured to obtain an underwater turtle image; a feature capturing module is configured to construct an image encoder based on a weight low-rank decomposition and updating mechanism, specifically, designing a weight low-rank decomposition and updating mechanism, performing lightweight and efficient fine-tuning on the encoder, introducing a low-rank parameterization method, decomposing part of the weights of the pre-trained model into low-rank matrices, and fine-tuning the decomposed low-rank matrices to adapt to specific downstream tasks; capturing global features of the obtained turtle image according to the image encoder and a multi-head self-attention mechanism; a feature reinforcement module is configured to perform feature selective reinforcement on the obtained global features of the turtle image to obtain a multi-scale reconstructed feature map; An instance segmentation module configured to generate candidate regions on the obtained multi-scale reconstructed feature map, perform alignment of the multi-scale reconstructed feature map on the generated candidate regions, perform detection segmentation of the aligned multi-scale reconstructed feature map through a multi-task branch network, and complete the sea turtle instance segmentation.

8. A computer readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by the processor, implements the steps of the sea turtle instance segmentation method based on weight low-rank decomposition and feature enhancement according to any one of claims 1-6.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, The processor implements the steps of the sea turtle instance segmentation method based on weight low-rank decomposition and feature enhancement according to any one of claims 1-6 when executing the program.

10. A computer program product comprising software code, characterized in that, The program in the software code implements the steps of the sea turtle instance segmentation method based on weight low-rank decomposition and feature enhancement according to any one of claims 1-6.

Images