Image segmentation method and system based on multi-scale feature enhancement and prototype fusion

Abstract

The invention discloses an image segmentation method and system based on multi-scale feature enhancement and prototype fusion, and relates to the technical field of image processing, and the method comprises the steps: inputting a support image and a query image into a residual network Resnet, and respectively extracting support features and query features; inputting the support features and the query features into a multi-scale feature enhancement module to obtain enhanced support features and enhanced query features; inputting the enhanced support feature and the enhanced query feature into a prototype fusion module to obtain a final mask; and inputting the final mask into a segmentation head, and segmenting the final mask to obtain a segmentation result of the query image. According to the method, the multi-scale feature enhancement module and the prototype fusion module are combined, and the problems of large intra-class difference and small inter-class difference in small sample image semantic segmentation are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to an image segmentation method and system based on multi-scale feature enhancement and prototype fusion. BACKGROUND

[0002] Image segmentation is to extract the region needed in the image for subsequent processing such as target recognition, classification, etc.

[0003] The existing image semantic segmentation method mainly relies on a large number of labeled images for training, uses a deep neural network to extract high-level semantic information and realizes pixel-level classification, and performs well in standard tasks. However, these methods face two major challenges: large intra-class variance, due to the lack of reference images, the target appearance changes significantly, and it is difficult to form a unified representation; high inter-class similarity, different class features are similar, and it is easy to be confused in complex scenes due to background interference. Small sample image semantic segmentation usually adopts a meta-learning strategy, which calculates prototype features or feature matching through support sets and query sets for classification, but the existing fixed prototype or dense matching strategy still cannot effectively cope with the above challenges, limiting its wide application in real scenes. SUMMARY

[0004] The embodiments of the present application provide an image segmentation method and system based on multi-scale feature enhancement and prototype fusion, to solve the problems of large intra-class variance, high inter-class similarity and insufficient data labeling in small sample image semantic segmentation in the prior art.

[0005] In one aspect, the embodiments of the present application provide an image segmentation method based on multi-scale feature enhancement and prototype fusion, comprising: inputting a support image and a query image into a residual network Resnet to extract support features and query features, respectively; inputting the support features and the query features into a multi-scale feature enhancement module, the multi-scale feature enhancement module comprising a multi-feature fusion unit and a feature enhancement unit, the multi-feature fusion unit comprising a parallel serial convolution branch and a feature selection branch, the serial convolution branch extracting multi-scale spatial information of the support features and the query features through a plurality of serial convolution blocks, the feature selection branch modeling attention of key regions in the support features and the query features through spatial block division and a feedforward network, and the multi-scale spatial information and the attention modeling result being added to form support fusion features and query fusion features; the feature enhancement unit adaptively adjusting weights of the support fusion features and the query fusion features through a channel-level attention regulation mechanism to obtain enhanced support features and enhanced query features, respectively; The enhanced support feature and the enhanced query feature are input into a prototype fusion module. The prototype fusion module first calculates an initial prototype according to the enhanced support feature, then generates an initial segmentation mask according to the initial prototype and the enhanced query feature, and then updates the initial segmentation mask through multiple stages of iterative optimization to obtain a final mask. The final mask is input into a segmentation head, and the segmentation head performs segmentation on the final mask to obtain a segmentation result of the query image.

[0006] On the other hand, the embodiment of the present application also provides an image segmentation system based on multi-scale feature enhancement and prototype fusion, comprising: a residual network Resnet, used for extracting support features and query features from input support images and query images respectively; a multi-scale feature enhancement module, which includes a multi-feature fusion unit and a feature enhancement unit. The multi-feature fusion unit includes a parallel serial convolution branch and a feature selection branch. The serial convolution branch extracts multi-scale spatial information of the support features and the query features through multiple serial convolution blocks. The feature selection branch models attention of key regions in the support features and the query features through spatial block division and a feedforward network. The multi-scale spatial information and the attention modeling result are added to form support fusion features and query fusion features. The feature enhancement unit adaptively adjusts the weights of the support fusion features and the query fusion features through a channel-level attention regulation mechanism to obtain enhanced support features and enhanced query features respectively; a prototype fusion module, used for first calculating an initial prototype according to the enhanced support features, then generating an initial segmentation mask according to the initial prototype and the enhanced query features, and then updating the initial segmentation mask through multiple stages of iterative optimization to obtain a final mask; a segmentation head, used for performing segmentation on the final mask to obtain a segmentation result of the query image.

[0007] The image segmentation method and system based on multi-scale feature enhancement and prototype fusion in the present application have the following advantages: 1. Stronger generalization: dynamically fusing support set and query set prototypes can adapt to target appearance changes.

[0008] 2. Higher accuracy: the multi-scale feature enhancement (MFE) module significantly enhances feature representation and improves boundary detail recognition ability.

[0009] 3. Significant practical advantages: on the PASCAL- 5i (pattern analysis, statistical modeling and computational learning) and COCO- 20i (context common object) data sets, the mIoU (mean intersection over union) is 73.76% and 51.6% respectively, which is better than existing methods. BRIEF DESCRIPTION OF DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced. Obviously, the accompanying drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0011] Figure 1 The architecture diagram of the image segmentation method based on multi-scale feature enhancement and prototype fusion provided by the embodiments of the present application.

[0012] Figure 2 The architecture diagram of the multi-feature fusion unit provided by the embodiments of the present application.

[0013] Figure 3 The architecture diagram of the feature enhancement unit provided by the embodiments of the present application.

[0014] Figure 4 The architecture diagram of the prototype fusion module and the segmentation head provided by the embodiments of the present application.

[0015] Figure 5 The comparison diagram of the visualization results provided by the embodiments of the present application. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments only constitute some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0017] Figures 1-4 The architecture diagram of the image segmentation method based on multi-scale feature enhancement and prototype fusion provided by the embodiments of the present application. The embodiments of the present application provide an image segmentation method based on multi-scale feature enhancement and prototype fusion, which comprises: S100, inputting a support image I s and a query image I q Residual network Resnet is inputted to extract support features and query features respectively.

[0018] Exemplarily, after the support image and the query image are obtained, pre-processing will be performed first, including size standardization and pixel value normalization, to adapt to the input requirements of the model.

[0019] The residual network Resnet includes the first layer Layer 0, the second layer Layer 1, the third layer Layer 2 and the fourth layer Layer 3. The first layer Layer 0 extracts support features and query features.

[0020] S110, input the supporting features and query features into the multi-scale feature enhancement module, the multi-scale feature enhancement module includes a multi-feature fusion unit and a feature enhancement unit, the multi-feature fusion unit includes a parallel serial convolution branch and a feature selection branch, the serial convolution branch extracts the multi-scale spatial information of the supporting features and the query features through multiple serial convolution blocks, the feature selection branch performs attention modeling on the key areas in the supporting features and the query features through spatial blocking and a feedforward network, and the results of the multi-scale spatial information and the attention modeling are added to form the supporting fusion features and the query fusion features; the feature enhancement unit adaptively adjusts the weights of the supporting fusion features and the query fusion features through the channel-level attention regulation mechanism to obtain enhanced supporting features and enhanced query features, respectively.

[0021] Exemplarily, the second layer Layer 1 performs convolution processing on the support fusion features and query fusion features output by the multi-feature fusion unit, and inputs the result of the convolution processing into the first feature enhancement unit. The third layer Layer 2 performs convolution processing on the enhanced support features and enhanced query features output by the first feature enhancement unit, and inputs the result of the convolution processing into the second feature enhancement unit. The fourth layer Layer 3 performs convolution processing on the enhanced support features and enhanced query features output by the second feature enhancement unit, and inputs the result of the convolution processing into the prototype fusion module.

[0022] Furthermore, the weights of the first layer Layer 0 and the second layer Layer 1 remain frozen, and the third layer Layer 2 and the fourth layer Layer 3 are convolved according to the weights adaptively adjusted by the first feature enhancement unit and the second feature enhancement unit, respectively.

[0023] Furthermore, if Figure 2 As shown in the figure, in the serial convolution branch, three sequentially connected convolution kernel 3×3 convolution blocks are used to extract multi-scale spatial information; in the feature selection branch, the input features are sequentially subjected to point-by-point convolution (PWConv), expansion and reshaping averaging ( U & R & M ) processing, feedforward network (FFN) processing, activation (Softmax) and selection (selection) processing to complete the attention modeling of key areas.

[0024] Specifically, the selection process includes the following steps: cosine similarity calculation, clipping, and reshaping of bilinear difference ( R & I) and point-wise convolution. In the cosine similarity calculation, the activation result is compared with the learnable vector ( Prompt ) to calculate cosine similarity, and the result is clipped. Furthermore, the cosine similarity result is multiplied pixel by pixel with the clipped result, and then reshaped by bilinear interpolation.

[0025] After obtaining and adding the results of multi-scale spatial information and attention modeling, they will undergo random dropout (Dropout), batch normalization (BN) and ReLU activation processing in sequence to finally obtain the fused features.

[0026] Furthermore, if Figure 3 As shown, the feature enhancement unit performs global average pooling on the input features in sequence ( GAP ), convolution (Conv), stack (Stack), average (Mean), activation (Sigmoid) and expansion (Expand). The result of the expansion processing is multiplied pixel by pixel with the result of the global average pooling to form the enhanced feature.

[0027] Specifically, in the convolution processing, the results of the global average pooling will be convolved using convolution blocks with convolution kernels of 3×3 and 5×5 respectively, and the output results of the two convolution blocks will be stacked together.

[0028] The serial convolution branch in this application captures multi-scale spatial information, the feature selection branch realizes attention modeling of key areas through spatial partitioning and feedforward network, and the feature enhancement unit adaptively adjusts feature weights through the channel-level attention regulation mechanism to highlight task-related areas and suppress background interference or non-target information, thereby generating more discriminative feature expressions. These two units work together to significantly improve the ability of image features to focus on target areas, providing high-quality input for subsequent prototype construction and mask prediction. Experimental results show that the introduction of this module can effectively improve segmentation accuracy on different baseline models and significantly enhance the overall performance of small-sample image semantic segmentation systems.

[0029] S120, the enhanced support features and the enhanced query features are input into the prototype fusion module. The prototype fusion module first calculates the initial prototype based on the enhanced support features, and then generates an initial segmentation mask based on the initial prototype and the enhanced query features. Then, the initial segmentation mask is updated through multiple stages of iterative optimization to obtain the final mask.

[0030] For example, Figure 4 As shown, the prototype fusion module dynamically optimizes the prototype representation by leveraging the feature correspondences between support images and query images, overcoming the limitations of static prototypes in traditional methods.

[0031] In the embodiments of the present application, when calculating the initial prototypes and generating the initial segmentation mask, the prototype fusion module first performs mask average pooling on the enhanced support features using the support mask MAP , to obtain initial prototypes of the foreground and the background respectively, and then generates the initial segmentation mask based on the cosine similarity between the enhanced query features and the initial prototypes of the foreground and the background.

[0032] The initial prototypes are represented as:

[0033] wherein, and represent the features of the foreground and the background respectively, represents the mask average pooling processing, F represents the features extracted by the feature extraction network, M represents the mask of the support image, and represent the initial prototypes of the foreground and the background respectively, represents the average processing, represents the splicing processing.

[0034] The initial segmentation mask can be represented as:

[0035] wherein, represents the initial segmentation mask, represents the calculation of the cosine similarity, represents the enhanced query features.

[0036] Further, when iteratively optimizing and updating the initial segmentation mask, a one-stage query prototype is first generated based on the enhanced query features and the initial segmentation mask, the one-stage query prototype and the initial prototype are weighted to obtain a one-stage enhanced prototype, and a one-stage segmentation mask is generated based on the cosine similarity between the enhanced query features and the one-stage enhanced prototype; then a two-stage query prototype is generated based on the enhanced query features and the one-stage segmentation mask, the initial prototype, the one-stage enhanced prototype and the two-stage query prototype are weighted to obtain a two-stage enhanced prototype, and a two-stage segmentation mask is generated by weighting and combining the one-stage segmentation mask and the two-stage segmentation mask to enhance the contribution degree of the most relevant region; then a three-stage query prototype is generated based on the enhanced query features and the two-stage segmentation mask, the initial prototype, the one-stage enhanced prototype, the two-stage enhanced prototype and the three-stage query prototype are weighted to obtain a three-stage enhanced prototype, and a prediction mask is generated based on the cosine similarity between the enhanced query features and the three-stage enhanced prototype; and the one-stage segmentation mask, the two-stage segmentation mask and the prediction mask are weighted and summed to obtain a final mask.

[0037] The one-stage prototype is represented as:

[0038] where, and are one-stage prototypes for foreground and background, respectively, , , and are weights for each term, respectively, and are one-stage query prototypes for foreground and background, respectively.

[0039] Two-stage prototypes are represented as:

[0040] where, and are two-stage prototypes for foreground and background, respectively, and are weights for each term, respectively, and are two-stage query prototypes for foreground and background, respectively.

[0041] Two-stage segmentation masks can be represented as:

[0042] where, and are two-stage segmentation masks and one-stage segmentation masks, respectively, and are hyperparameters for controlling the contribution of support features and query features.

[0043] Three-stage prototypes are represented as:

[0044] where, and are two-stage prototypes for foreground and background, respectively, and are weights for each term, respectively, and are two-stage query prototypes for foreground and background, respectively.

[0045] Final masks are represented as:

[0046] where, is the final mask, is the predicted mask.

[0047] Through multi-stage iterative optimization of the prototype, the prototype fusion module enhances the category discrimination ability, adapts to specific features of the query image, and dynamically optimizes the prototype and the multi-stage supervised training strategy, which significantly improves the segmentation performance, especially in complex background and diversified object morphology, and exhibits good adaptability and robustness.

[0048] The above hyperparameters and the weights of each item are obtained through training of the prototype fusion module. In the training process of the prototype fusion module, the loss function adopts cross-entropy loss, which is represented as:

[0049] wherein, represents the loss function adopted in the training of the prototype fusion module, represents the cross-entropy loss, represents the real mask of the query image in the training set. The loss function can ensure consistent learning in the iteration process.

[0050] In the training of the multi-scale feature enhancement module, the loss function also adopts cross-entropy loss, which is represented as:

[0051] wherein, represents the loss function adopted in the training of the multi-scale feature enhancement module, represents the prototype generated by the enhanced query feature generated by the multi-scale feature enhancement module. The loss function can ensure the robustness of the network in the feature extraction and prototype updating process. In the training process, the samples in the support set and the query set also need to be preprocessed before being input into the model, and the model also needs to be initialized with network parameters (including the weights of the convolution block, the parameters of the attention mechanism and the hyperparameters of the optimizer) before starting the training.

[0052] Therefore, the loss function for model training in the present application is represented as:

[0053] The loss function optimizes the performance of the model in the small sample scene through the supervision of multi-stage prediction and the self-supervised signal of the query feature, and ensures the consistency and accuracy of the segmentation result under complex background and diversified object appearance.

[0054] S130, input the final mask into the segmentation head, and the segmentation head segments the final mask to obtain the segmentation result of the query image.

[0055] The present application also provides an image segmentation system based on multi-scale feature enhancement and prototype fusion, comprising: Residual network Resnet, used for extracting support features and query features from input support images and query images respectively; The multi-scale feature enhancement module includes a multi-feature fusion unit and a feature enhancement unit. The multi-feature fusion unit includes a parallel serial convolution branch and a feature selection branch. The serial convolution branch extracts multi-scale spatial information of the support features and the query features through a plurality of serially connected convolution blocks. The feature selection branch models attention of key regions in the support features and the query features through spatial partitioning and a feedforward network. The multi-scale spatial information and the attention modeling results are added to form support fusion features and query fusion features. The feature enhancement unit adaptively adjusts weights of the support fusion features and the query fusion features through a channel-level attention regulation mechanism to obtain enhanced support features and enhanced query features. The prototype fusion module is used to calculate initial prototypes according to the enhanced support features, generate initial segmentation masks according to the initial prototypes and the enhanced query features, and then update the initial segmentation masks through a plurality of stages of iterative optimization to obtain final masks. The segmentation head is used to segment the final masks to obtain a segmentation result of the query image.

[0056] Experimental verification The experiment was performed under the PyTorch framework using four NVIDIA GeForce GTX 3090 GPUs. During the model training process, a stochastic gradient descent (SGD) optimizer was used, with a momentum of 0.9, a weight decay coefficient of 5e-4, and an initial learning rate of 0.001. When Resnet50 was used as the backbone network, each training batch contained 16 pairs of support-query images. When Resnet101 was used, each batch contained 8 image pairs. The PASCAL- 5i and COCO- 20i datasets were evaluated under a 5-sample setting, using a Resnet101 backbone network, and mIoU scores of 73.76 and 51.6 were obtained, as shown in Table 1.

[0057] Table 1. Quantitative comparison of the method of the present application with the optimal model

[0058] Rife (Relevant Intrinsic Feature Enhancement Network), DCP (Divide-and-Conquer Proxies), DRNet (Disentanglement and Recombination Network), and DBM (Dual Branch Multi-Level Network) are the best models for small sample segmentation performance in 2024. Rife is proposed in the literature “X. Bao, J. Qin, S. Sun, X. Wang, and Y. Zheng. Relevant intrinsic feature enhancement network for few-shot semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence , 2024.” DCP is proposed in the literature “C. Lang, G. Cheng, B. Tu, and J. Han. Few-shot segmentation via divide-and-conquer proxies. International Journal of Computer Vision IEEE Transactions on , 2024.” DRNet is proposed in the literature “Z. Chang, X. Gao, N. Li, H. Zhou, and Y. Lu. Drnet: Disentanglement and recombination network for few-shot semantic segmentation. Circuits and Systems for Video Technology IEEE Transactions on Image , 2024.” DBM is proposed in the literature “Y. Chen, R. Jiang, Y. Zheng, B. Sheng, Z.-X. Yang, and E. Wu. Dual branch multi-level semantic learning for few-shot segmentation. Processing

[0059] 1-shot and 5-shot represent the experimental results when the number of labeled samples is 1 and 5, respectively. Fold-0, Fold-1, Fold-2, and Fold-3 represent the division of the dataset into four subsets according to the class. In the 5-shot case, the proposed method outperforms existing methods in the mIoU evaluation indicator, demonstrating the superiority of the proposed method.

[0060] As shown in Figure 5 , the visualization results of the comparison between the method of the present application and the baseline method, which does not use the multi-scale feature enhancement module in the present application and lacks a stage of prototype fusion module, and the rest of the structure is the same as the present application.​Figure 5 The middle real mask and the segmentation result are represented by red areas, and the visual results clearly reflect that the method proposed in the application has better effects in target segmentation and background consistency.

[0061] Although preferred embodiments of the application have been described, those skilled in the art will be able to make additional changes and modifications without departing from the spirit and scope of the application. Therefore, the appended claims are intended to cover all such changes and modifications that fall within the scope of the application.

[0062] Obviously, various modifications and changes can be made to the present application by those skilled in the art without departing from the spirit and scope of the present application. Thus, it is intended that the present application embrace all such changes and modifications as fall within the scope of the appended claims and their equivalents.

Claims

1. An image segmentation method based on multi-scale feature enhancement and prototype fusion, characterized in that, The method comprises the following steps: inputting a support image and a query image into a residual network Resnet to extract support features and query features respectively; inputting the support features and the query features into a multi-scale feature enhancement module, the multi-scale feature enhancement module comprising a multi-feature fusion unit and a feature enhancement unit, the multi-feature fusion unit comprising a parallel serial convolution branch and a feature selection branch, the serial convolution branch extracting multi-scale spatial information of the support features and the query features through a plurality of convolution blocks connected in series, the feature selection branch modeling attention of key regions in the support features and the query features through spatial division and a feedforward network, and the multi-scale spatial information and the attention modeling result being added to form support fusion features and query fusion features; the feature enhancement unit adaptively adjusting weights of the support fusion features and the query fusion features through a channel-level attention regulation mechanism to obtain enhanced support features and enhanced query features respectively; inputting the enhanced support features and the enhanced query features into a prototype fusion module, the prototype fusion module first calculating initial prototypes according to the enhanced support features, then generating an initial segmentation mask according to the initial prototypes and the enhanced query features, and then updating the initial segmentation mask through a plurality of stages of iterative optimization to obtain a final mask; inputting the final mask into a segmentation head, the segmentation head performing segmentation on the final mask to obtain a segmentation result of the query image.

2. The image segmentation method based on multi-scale feature enhancement and prototype fusion according to claim 1, characterized in that, The residual network Resnet comprises a first layer Layer 0, a second layer Layer 1, a third layer Layer 2 and a fourth layer Layer 3, the first layer Layer 0 extracts the support features and the query features, the second layer Layer 1 performs convolution processing on the support fusion features and the query fusion features output by the multi-feature fusion unit, inputs the convolution processing result into a first feature enhancement unit, the third layer Layer 2 performs convolution processing on the enhanced support features and the enhanced query features output by the first feature enhancement unit, inputs the convolution processing result into a second feature enhancement unit, and the fourth layer Layer 3 performs convolution processing on the enhanced support features and the enhanced query features output by the second feature enhancement unit, and inputs the convolution processing result into the prototype fusion module. 3.The image segmentation method based on multi-scale feature enhancement and prototype fusion according to claim 2, characterized in that, The weights of the first layer Layer 0 and the second layer Layer 1 remain frozen, and the third layer Layer 2 and the fourth layer Layer 3 perform convolution processing according to the weights adaptively adjusted by the first feature enhancement unit and the second feature enhancement unit respectively.

4. The image segmentation method based on multi-scale feature enhancement and prototype fusion according to claim 1, characterized in that, In the serial convolution branch, three convolution blocks with a convolution kernel of 3×3 connected in sequence are used to extract the multi-scale spatial information; in the feature selection branch, point-by-point convolution, unfolding and reshaping average processing, feedforward network processing, activation and selection processing are sequentially performed on the input features to complete attention modeling of key regions.

5. The image segmentation method based on multi-scale feature enhancement and prototype fusion according to claim 4, characterized in that, The selection processing includes cosine similarity calculation, remolding bilinear difference and point-by-point convolution in sequence.

6. The image segmentation method based on multi-scale feature enhancement and prototype fusion according to claim 1, characterized in that, The feature enhancement unit sequentially performs global average pooling, convolution, stacking, averaging, activation and expansion on the input features, pixel-by-pixel multiplication is performed between the results of the expansion processing and the results of the global average pooling to form the enhanced features.

7. The image segmentation method based on multi-scale feature enhancement and prototype fusion according to claim 1, characterized in that, In calculating the initial prototype and generating the initial segmentation mask, the prototype fusion module first performs mask average pooling on the enhanced support features using a support mask to obtain the initial prototype of the foreground and the background, respectively, and then generates the initial segmentation mask based on the cosine similarity between the enhanced query features and the initial prototypes of the foreground and the background.

8. The image segmentation method based on multi-scale feature enhancement and prototype fusion according to claim 1, characterized in that, In the iterative optimization and update of the initial segmentation mask, a one-stage query prototype is first generated based on the enhanced query features and the initial segmentation mask, the one-stage query prototype and the initial prototype are weighted to obtain a one-stage enhanced prototype, and a one-stage segmentation mask is generated based on the cosine similarity between the enhanced query features and the one-stage enhanced prototype; Then, a two-stage query prototype is generated based on the enhanced query features and the one-stage segmentation mask, the initial prototype, the one-stage enhanced prototype and the two-stage query prototype are weighted to obtain a two-stage enhanced prototype, and a two-stage segmentation mask is generated by weighting and combining the one-stage segmentation mask and the two-stage segmentation mask; then, a three-stage query prototype is generated based on the enhanced query features and the two-stage segmentation mask, the initial prototype, the one-stage enhanced prototype, the two-stage enhanced prototype and the three-stage query prototype are weighted to obtain a three-stage enhanced prototype, and a predicted mask is generated based on the cosine similarity between the enhanced query features and the three-stage enhanced prototype; The one-stage segmentation mask, the two-stage segmentation mask and the predicted mask are weighted and summed to obtain the final mask.

9. The system for image segmentation based on multi-scale feature enhancement and prototype fusion according to any one of claims 1-8, characterized in that, It comprises: a residual network Resnet for extracting support features and query features from input support images and query images, respectively; a multi-scale feature enhancement module, which comprises a multi-feature fusion unit and a feature enhancement unit, the multi-feature fusion unit comprising a serial convolution branch and a feature selection branch in parallel, the serial convolution branch extracting multi-scale spatial information of the support features and the query features through a plurality of serial convolution blocks, and the feature selection branch modeling attention of key regions in the support features and the query features through spatial block division and a feedforward network, the multi-scale spatial information and the attention modeling results being added to form support fusion features and query fusion features; the feature enhancement unit adaptively adjusting weights of the support fusion features and the query fusion features through a channel-level attention regulation mechanism to obtain enhanced support features and enhanced query features, respectively; a prototype fusion module for calculating an initial prototype according to the enhanced support features, generating an initial segmentation mask according to the initial prototype and the enhanced query features, and then iteratively optimizing and updating the initial segmentation mask through multiple stages to obtain a final mask; and a prototype fusion module for calculating an initial prototype according to the enhanced support features, generating an initial segmentation mask according to the initial prototype and the enhanced query features, and then iteratively optimizing and updating the initial segmentation mask through multiple stages to obtain a final mask. The segmentation head is configured to segment the final mask to obtain a segmentation result of the query image.

Images