CN113724276B - Polyp image segmentation method and device - Google Patents

Abstract

The embodiment of the invention discloses a polyp image segmentation method and device, comprising the following steps: acquiring a polyp image to be input; selecting a reference image with a color different from that of the polyp image from a preset training set, and exchanging the reference image with the color of the polyp image; extracting shallow features and deep features from the color-exchanged polyp image, inhibiting background noise of the shallow features by using a shallow attention model, and fusing the shallow features and the deep features; and (3) carrying out predictive response value re-equalization processing on the fused features by adopting a probability correction strategy model to obtain polyp feature images with clear edges. The invention can accurately and efficiently divide the polyp part from the image and has better generalization in various complex actual scenes.

Description

A method and device for segmenting polyp images

Technical Field

The embodiments of the present invention relate to the field of image processing technology, and in particular to a method and device for segmenting a polyp image.

Background technique

Polyps are prone to cancer, especially multiple polyps, so early screening and treatment of polyps is very necessary. Polyp segmentation, as a computer vision task, can automatically segment polyps in images or videos, greatly reducing the workload of doctors. Therefore, establishing an accurate polyp segmentation model is of great significance for clinical medical diagnosis.

At present, PraNet based on parallel reverse attention network is the most commonly used existing technology. PraNet first uses Res2Net neural network to extract features of different semantic levels from polyp images, and then uses parallel decoders to aggregate high-semantic-level features to obtain the global context information of the image. However, since the high-semantic-level features lose too much detail information, the polyp segmentation results obtained are relatively rough. In order to further explore the polyp boundary clues, PraNet uses the reverse attention module to construct the relationship between the polyp area and the polyp boundary. Through the continuous interaction and complementation between the polyp area and the polyp boundary, PraNet can obtain more accurate polyp segmentation prediction results.

Although PraNet can obtain relatively accurate results, it has two important defects: (1) The segmentation results of small polyps are poor. This is because small polyps lose too much information in high-level semantic features and are difficult to recover directly; in addition, there are large errors in the annotation of small polyp boundaries, which has a great impact on the final segmentation results; (2) It ignores the color bias in the dataset. Generally, there are large differences in the colors of polyp images collected under different conditions. This difference will interfere with the training of the polyp segmentation model. Especially when there are fewer training images, the model can easily overfit to the polyp color, resulting in a significant decrease in the generalization ability of the model in actual application scenarios.

Summary of the invention

In order to solve the above technical problems, an embodiment of the present invention provides a polyp image segmentation method, comprising:

Acquire a polyp image to be input;

Selecting a reference image having a different color from the polyp image from a preset training set, and exchanging the colors of the reference image and the polyp image;

Extracting shallow features and deep features from the color-swapped polyp image, suppressing background noise of the shallow features using a shallow attention model, and fusing the shallow features with the deep features;

The probability correction strategy model is used to re-balance the predicted response values of the fused features to obtain a polyp feature image with clear edges.

Furthermore, exchanging the colors of the reference image and the polyp image includes:

Convert the colors of the polyp image X1 and the reference image X2 from the RGB color space to the LAB color space to obtain color values L1 and L2 of the polyp image X1 and the reference image X2 in the LAB color space;

Calculate the mean and standard deviation of the channels of the polyp image X1 in the LAB color space and the mean and standard deviation of the channels of the reference image X2 in the LAB color space;

The color value of the polyp image Y1 in the RGB color space and the color value of the reference image Y2 in the RGB color space are obtained using a preset color conversion formula.

Furthermore, the method of suppressing the background noise of the shallow features by using the shallow attention model includes:

Upsampling the deep features by bilinear interpolation so that the resolution of the sampled deep features is the same as that of the shallow features;

Selecting elements greater than 0 from the sampled deep features as the attention map of the shallow features to obtain the deep features to be fused;

The deep features to be fused are multiplied element by element with the shallow features to obtain the shallow features after suppressing the background noise.

Furthermore, the fusing of the shallow features and the deep features includes:

Extracting the first, second and third features of the last three scales when the shallow features after suppressing background noise are processed by a convolutional neural network;

Fusing the first feature and the second feature to obtain a first fused feature;

Fusing the second feature and the third feature to obtain a second fused feature;

The first fusion feature and the second fusion feature are spliced according to channels to obtain a final fusion feature.

Furthermore, the probability correction strategy model is used to rebalance the predicted response values of the fused features to obtain a polyp feature image with clear edges, including:

Counting the number of pixels whose characteristic response values are greater than 0 in the fused polyp characteristic image to obtain a first pixel value;

Counting the number of pixels whose characteristic response values are less than 0 in the fused polyp characteristic image to obtain a second pixel value;

The first pixel value and the second pixel value are normalized, and the characteristic response value greater than 0 in the polyp characteristic image is divided by the normalized first pixel value, and the characteristic response value less than 0 in the polyp characteristic image is divided by the normalized second pixel value to obtain a corrected polyp characteristic image.

A polyp image segmentation device, comprising:

An acquisition module, used for acquiring a polyp image to be input;

A processing module, configured to select a reference image having a different color from the polyp image from a preset training set, and exchange the colors of the reference image with those of the polyp image;

The processing module is used to extract shallow features and deep features from the polyp image after color exchange, suppress background noise of the shallow features by using a shallow attention model, and fuse the shallow features with the deep features;

The execution module is used to use the probability correction strategy model to re-balance the predicted response values of the fused features to obtain a polyp feature image with clear edges.

Furthermore, the processing module includes:

A first processing submodule is used to convert the colors of the polyp image X1 and the reference image X2 from the RGB color space to the LAB color space to obtain color values L1 and L2 of the polyp image X1 and the reference image X2 in the LAB color space;

A second processing submodule, used for calculating the mean and standard deviation of the channels of the polyp image X1 in the LAB color space and the mean and standard deviation of the channels of the reference image X2 in the LAB color space;

The third processing submodule is used to obtain the color value of the polyp image Y1 in the RGB color space and the color value of the reference image Y2 in the RGB color space by using a preset color conversion formula.

Furthermore, the processing module includes:

A fourth processing submodule, configured to upsample the deep features by bilinear interpolation so that the resolution of the sampled deep features is the same as that of the shallow features;

The first acquisition submodule is used to select elements greater than 0 from the sampled deep features to determine as the attention map of the shallow features, so as to obtain the deep features to be fused;

The first execution submodule is used to multiply the deep features to be fused and the shallow features element by element to obtain the shallow features after suppressing background noise.

Furthermore, the processing module includes:

The second acquisition submodule is used to extract the first feature, the second feature and the third feature of the last three scales when the shallow features after suppressing the background noise are processed by the convolutional neural network;

a fifth processing submodule, configured to fuse the first feature and the second feature to obtain a first fused feature;

A sixth processing submodule, configured to fuse the second feature and the third feature to obtain a second fused feature;

The second execution submodule is used to splice the first fusion feature and the second fusion feature according to channels to obtain a final fusion feature.

Furthermore, the execution module includes:

The third acquisition submodule is used to count the number of pixels whose characteristic response values are greater than 0 in the fused polyp characteristic image to obtain a first pixel value;

A fourth acquisition submodule is used to count the number of pixels whose characteristic response values are less than 0 in the fused polyp characteristic image to obtain a second pixel value;

The third execution submodule is used to normalize the first pixel value and the second pixel value, and divide the characteristic response value greater than 0 in the polyp characteristic image by the normalized first pixel value, and divide the characteristic response value less than 0 in the polyp characteristic image by the normalized second pixel value to obtain a corrected polyp characteristic image.

The beneficial effects of the embodiments of the present invention are:

(1) To address the problem of inaccurate segmentation of small polyp targets, the shallow attention module (SAM) in the present invention can enhance the model's ability to extract and utilize shallow features of the neural network, because shallow features retain more detailed features for small polyps. Unlike the traditional method of directly fusing multiple layers of features through operations such as addition or splicing, SAM uses deep features as an auxiliary and guides the removal of background noise in shallow features through the attention mechanism, thereby greatly improving the usability of shallow features. In addition, the foreground and background pixel distributions of small polyp images are unbalanced. To this end, the probability correction strategy (PCS) can be used to dynamically and adaptively correct its response value according to the prediction results during the model inference stage, thereby optimizing the edge of the segmented target and reducing the impact of the imbalanced foreground and background distribution.

(2) In response to the color deviation problem of the dataset, the present invention proposes a color exchange (CE) operation to eliminate the impact of color deviation on model training. Specifically, through CE, the colors of different images can be transferred to each other, and the colors of the same image can change differently, thereby achieving the decoupling of image color and image content. Therefore, the model can focus on the image content itself during training without being disturbed by its color. A large number of quantitative and qualitative experiments show that the SANet model proposed in the present invention can accurately and efficiently segment polyp parts from images and has better generalization in various complex practical scenarios.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

FIG1 is a schematic flow chart of a polyp image segmentation method provided by an embodiment of the present invention;

FIG2 is a schematic diagram of effect comparison provided by an embodiment of the present invention;

FIG3 is another schematic diagram of effect comparison provided by an embodiment of the present invention;

FIG4 is a schematic diagram of the structure of a polyp image segmentation device provided by an embodiment of the present invention;

FIG5 is a basic structural block diagram of a computer device provided in an embodiment of the present invention.

Detailed ways

In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

Please refer to FIG. 1 , which is a polyp image segmentation method provided by an embodiment of the present invention, including:

S1100, obtaining a polyp image to be input;

S1200, selecting a reference image having a different color from the polyp image from a preset training set, and exchanging the colors of the reference image and the polyp image;

S1300, extracting shallow features and deep features from the color-swapped polyp image, suppressing background noise of the shallow features using a shallow attention model, and fusing the shallow features with the deep features;

S1400. A probability correction strategy model is used to rebalance the predicted response values of the fused features to obtain a polyp feature image with clear edges.

The shallow attention module (SAM) in the present invention can enhance the model's ability to extract and utilize shallow features of neural networks, because shallow features retain more detailed features for small polyps. Different from the traditional method of directly fusing multiple layers of features through operations such as addition or splicing, SAM uses deep features as an auxiliary and guides the removal of background noise in shallow features through the attention mechanism, thereby greatly improving the availability of shallow features. In addition, the foreground and background pixel distribution of small polyp images is unbalanced. To this end, through the probability correction strategy (PCS), the response value can be dynamically and adaptively corrected according to the prediction results during the model reasoning stage, thereby optimizing the edge of the segmented target and reducing the impact of the imbalanced distribution of the foreground and background. In addition, in response to the color deviation problem of the data set, a color exchange (CE) operation proposed in the present invention is used to eliminate the impact of color deviation on model training. Specifically, through CE, the colors of different images can be migrated to each other, and the colors of the same image can change differently, thereby achieving the decoupling of image color and image content, so that the model can focus on the image content itself during training without being disturbed by its color. A large number of quantitative and qualitative experiments show that the SANet model proposed in the present invention can accurately and efficiently segment polyp parts from images and has better generalization in various complex practical scenarios.

The embodiment of the present invention includes three models CE, SAM and PCS, among which CE is used in the data augmentation stage to migrate the colors of different images to the input image; SAM is used in the feature fusion stage to give full play to the potential of shallow features; PCS is used in the model inference stage to optimize and adjust the prediction results.

Specifically, the color swap operation directly acts on the input image, so that the same input image can present different color styles during the model training process, as shown in Figure 2. Specifically, for any input image, a different color image is randomly selected from the training set as a reference and its color is transferred to the input image. Since the reference image selected each time is random, the same input image can present different color styles but the corresponding label remains unchanged. Therefore, the model will focus on the image content during training and will not be affected by the image color. Among them, exchanging the colors of the reference image and the polyp image includes:

Step 1: Convert the colors of the polyp image X1 and the reference image X2 from the RGB color space to the LAB color space to obtain the color values L1 and L2 of the polyp image X1 and the reference image X2 in the LAB color space;

Step 2: Calculate the mean and standard deviation of the channels of the polyp image X1 in the LAB color space and the mean and standard deviation of the channels of the reference image X2 in the LAB color space;

Step 3: Use a preset color conversion formula to obtain the color value of the polyp image Y1 in the RGB color space and the color value of the reference image Y2 in the RGB color space.

In the embodiment of the present invention, small polyp images face serious information loss problems in feature downsampling. Therefore, making full use of shallow features containing rich details is of great significance for small polyp target segmentation. However, due to the limitation of the receptive field, these features contain a large amount of background noise. For this reason, the present invention proposes that SAM uses deep features to suppress background noise of shallow features, which can fully improve the ease of use of shallow features and promote the segmentation effect of the model on small polyp targets. Specifically, the background noise of the shallow features is suppressed by using a shallow attention model, including:

Step 1: upsampling the deep features by bilinear interpolation so that the resolution of the sampled deep features is the same as that of the shallow features;

Step 2: Select elements greater than 0 from the sampled deep features to determine them as the attention map of the shallow features, and obtain the deep features to be fused;

Step 3: multiply the deep features to be fused and the shallow features element by element to obtain the shallow features after suppressing the background noise.

In one embodiment of the present invention, upsample to the same resolution size as and through bilinear interpolation, that is, set elements less than 0 to 0, as the attention map of , that is; and multiply and element-by-element to suppress background noise, that is.

In the embodiment of the present invention, SAM can effectively fuse the deep and shallow features together. The fusion of the shallow features and the deep features includes:

Extracting the first, second and third features of the last three scales when the shallow features after suppressing background noise are processed by a convolutional neural network;

Fusing the first feature and the second feature to obtain a first fused feature;

Fusing the second feature and the third feature to obtain a second fused feature;

The first fusion feature and the second fusion feature are spliced according to channels to obtain a final fusion feature.

In one embodiment of the present invention, in the SANet model, we fuse the features output by stage3, stage4 and stage5 of Res2Net (respectively denoted as f3, f4, f5) to reduce the amount of model calculation. Based on SAM, they will be fused together to give full play to the advantages of each scale feature. Fusion will be obtained; fusion will be obtained; splicing the newly obtained by channel will be obtained to obtain the final fusion feature.

In the embodiment of the present invention, there is a serious phenomenon of uneven distribution of foreground and background pixels in the small polyp image. Negative samples (background pixels) dominate the model training process. This prior bias causes the model to be more inclined to give positive samples (foreground pixels) lower response values (logit), resulting in poor target edge segmentation. In order to correct this imbalance, the present invention uses PCS to rebalance the predicted response values in the model inference stage, wherein the probability correction strategy model is used to rebalance the predicted response values of the fused features to obtain a polyp feature image with clear edges, including:

Counting the number of pixels whose characteristic response values are greater than 0 in the fused polyp characteristic image to obtain a first pixel value;

Counting the number of pixels whose characteristic response values are less than 0 in the fused polyp characteristic image to obtain a second pixel value;

The first pixel value and the second pixel value are normalized, and the characteristic response value greater than 0 in the polyp characteristic image is divided by the normalized first pixel value, and the characteristic response value less than 0 in the polyp characteristic image is divided by the normalized second pixel value to obtain a corrected polyp characteristic image.

In one embodiment of the present invention, the number of pixels with a response value greater than 0 (logit>0) is counted to obtain; the number of pixels with a response value less than 0 (logit<0) is counted to obtain; normalization is performed on, that is, the response value of logit>0 is divided by, and the response value of logit<0 is divided by finally obtaining the corrected polyp characteristic image. After PCS, the deviation of the prediction result caused by the number of positive and negative samples is eliminated, and a clearer prediction result can be obtained for the edge of the target. Figure 2 shows some details of the results obtained using PCS.

Table 1. Quantitative results of different models on the dataset

In one embodiment of the present invention, Table 1 shows the quantitative results of different models on five datasets, namely, Kvasir, CVC-ClinicDB, CVC-ColonDB, EndoScene, and ETIS. It can be seen that the present invention has achieved the highest score on all datasets. FIG3 shows the qualitative experimental results of different algorithms on specific images. It can be seen that the present invention can obtain a more complete and clearer polyp area than the previous model. Based on the above experiments, the present invention can better remove the deviation and background noise in the dataset, thereby having excellent performance in polyp segmentation.

As shown in FIG4 , in order to solve the above problems, an embodiment of the present invention further provides a polyp image segmentation device, comprising: an acquisition module 2100, a processing module 2200 and an execution module 2300, wherein the acquisition module 2100 is used to acquire a polyp image to be input; the processing module 2200 is used to select a reference image having a different color from the polyp image from a preset training set, and exchange the color of the reference image with that of the polyp image; the processing module 2200 is used to extract shallow features and deep features from the polyp image after the color exchange, and use a shallow attention model to suppress the background noise of the shallow features, and fuse the shallow features with the deep features; the execution module 2300 is used to use a probability correction strategy model to perform predicted response value re-equalization processing on the fused features to obtain a polyp feature image with clear edges.

In some embodiments, the processing module includes: a first processing submodule, used to convert the colors of the polyp image X1 and the reference image X2 from the RGB color space to the LAB color space, and obtain the color values L1 and L2 of the polyp image X1 and the reference image X2 in the LAB color space; a second processing submodule, used to calculate the mean and standard deviation of the channels of the polyp image X1 in the LAB color space and the mean and standard deviation of the channels of the reference image X2 in the LAB color space; a third processing submodule, used to use a preset color conversion formula to obtain the color value of the polyp image Y1 in the RGB color space and the color value of the reference image Y2 in the RGB color space.

In some embodiments, the processing module includes: a fourth processing sub-module, used to upsample the deep features through bilinear difference so that the resolution of the sampled deep features is the same as that of the shallow features; a first acquisition sub-module, used to select elements greater than 0 from the sampled deep features as the attention map of the shallow features, and obtain the deep features to be fused; a first execution sub-module, used to multiply the deep features to be fused and the shallow features element by element to obtain the shallow features after suppressing background noise.

In some embodiments, the processing module includes: a second acquisition submodule, used to extract the first feature, the second feature and the third feature of the last three scales when the shallow features after suppressing background noise are processed using a convolutional neural network; a fifth processing submodule, used to fuse the first feature and the second to obtain a first fused feature; a sixth processing submodule, used to fuse the second feature and the third feature to obtain a second fused feature; and a second execution submodule, used to splice the first fused feature and the second fused feature according to channels to obtain a final fused feature.

In some embodiments, the execution module includes: a third acquisition submodule, used to count the number of pixels with characteristic response values greater than 0 in the fused polyp characteristic image to obtain a first pixel value; a fourth acquisition submodule, used to count the number of pixels with characteristic response values less than 0 in the fused polyp characteristic image to obtain a second pixel value; a third execution submodule, used to normalize the first pixel value and the second pixel value, and divide the characteristic response values greater than 0 in the polyp characteristic image by the normalized first pixel value, and divide the characteristic response values less than 0 in the polyp characteristic image by the normalized second pixel value to obtain a corrected polyp characteristic image.

To solve the above technical problems, the embodiment of the present invention further provides a computer device. Please refer to FIG5 for details, which is a basic structural block diagram of the computer device of the present embodiment.

As shown in Figure 5, a schematic diagram of the internal structure of a computer device. As shown in Figure 5, the computer device includes a processor, a non-volatile storage medium, a memory, and a network interface connected via a system bus. Among them, the non-volatile storage medium of the computer device stores an operating system, a database, and a computer-readable instruction, and a control information sequence may be stored in the database. When the computer-readable instruction is executed by the processor, the processor can implement an image processing method. The processor of the computer device is used to provide computing and control capabilities to support the operation of the entire computer device. The memory of the computer device may store computer-readable instructions, and when the computer-readable instruction is executed by the processor, the processor can execute an image processing method. The network interface of the computer device is used to connect and communicate with the terminal. It can be understood by those skilled in the art that the structure shown in Figure 5 is only a block diagram of a partial structure related to the present application scheme, and does not constitute a limitation on the computer device to which the present application scheme is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

In this embodiment, the processor is used to execute the specific contents of the acquisition module 2100, the processing module 2200 and the execution module 2300 in FIG4, and the memory stores the program code and various data required to execute the above modules. The network interface is used to transmit data between the user terminal or the server. The memory in this embodiment stores the program code and data required to execute all submodules in the image processing method, and the server can call the program code and data of the server to execute the functions of all submodules.

In the computer device provided by the embodiment of the present invention, the reference feature map is obtained by extracting features from a high-definition image set in a reference pool. Due to the diversity of images in the high-definition image set, the reference feature map contains all possible local features, and can provide high-frequency texture information for each low-resolution image, which not only ensures the richness of features, but also reduces the memory burden. In addition, the reference feature map is searched based on the low-resolution image, and the selected reference feature map can adaptively shield or enhance a variety of different features, making the details of the low-resolution image richer.

The present invention also provides a storage medium storing computer-readable instructions, and when the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the image processing method described in any of the above embodiments.

Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

The above descriptions are only some embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims (6)

1. A method for segmenting a polyp image, comprising: Acquire a polyp image to be input; Selecting a reference image having a different color from the polyp image from a preset training set, and exchanging the colors of the reference image and the polyp image; Extracting shallow features and deep features from the color-swapped polyp image, suppressing background noise of the shallow features using a shallow attention model, and fusing the shallow features with the deep features; The probability correction strategy model is used to rebalance the predicted response values of the fused features to obtain a polyp feature image with clear edges. The method of suppressing the background noise of the shallow features by using the shallow attention model includes: Upsampling the deep features by bilinear interpolation so that the resolution of the sampled deep features is the same as that of the shallow features; Selecting elements greater than 0 from the sampled deep features as the attention map of the shallow features to obtain the deep features to be fused; Multiplying the deep features to be fused by the shallow features element by element to obtain shallow features after suppressing background noise; The probability correction strategy model is used to rebalance the predicted response values of the fused features to obtain a polyp feature image with clear edges, including: Counting the number of pixels whose characteristic response values are greater than 0 in the fused polyp characteristic image to obtain a first pixel value; Counting the number of pixels whose characteristic response values are less than 0 in the fused polyp characteristic image to obtain a second pixel value; The first pixel value and the second pixel value are normalized, and the characteristic response value greater than 0 in the polyp characteristic image is divided by the normalized first pixel value, and the characteristic response value less than 0 in the polyp characteristic image is divided by the normalized second pixel value to obtain a corrected polyp characteristic image. 2. The segmentation method according to claim 1, characterized in that exchanging the colors of the reference image and the polyp image comprises: Convert the colors of the polyp image X1 and the reference image X2 from the RGB color space to the LAB color space to obtain color values L1 and L2 of the polyp image X1 and the reference image X2 in the LAB color space; Calculate the mean of the channels of the polyp image X1 in the LAB color space and standard deviation/> And the mean value of the channel of the reference image X2 in the LAB color space/> and standard deviation/> ; The color value of the polyp image Y1 in the RGB color space and the color value of the reference image Y2 in the RGB color space are obtained using a preset color conversion formula. 3. The segmentation method according to claim 1, characterized in that the fusing of the shallow features and the deep features comprises: Extracting the first, second and third features of the last three scales when the shallow features after suppressing background noise are processed by a convolutional neural network; Fusing the first feature and the second feature to obtain a first fused feature; Fusing the second feature and the third feature to obtain a second fused feature; The first fusion feature and the second fusion feature are concatenated according to channels to obtain a final fusion feature. 4. A polyp image segmentation device, comprising: An acquisition module, used for acquiring a polyp image to be input; A processing module, configured to select a reference image having a different color from the polyp image from a preset training set, and exchange the colors of the reference image with those of the polyp image; The processing module is used to extract shallow features and deep features from the polyp image after color exchange, suppress background noise of the shallow features by using a shallow attention model, and fuse the shallow features with the deep features; An execution module is used to use a probability correction strategy model to perform predicted response value re-balancing processing on the fused features to obtain a polyp feature image with clear edges; Wherein, the processing module includes: A fourth processing submodule, configured to upsample the deep features by bilinear interpolation so that the resolution of the sampled deep features is the same as that of the shallow features; The first acquisition submodule is used to select elements greater than 0 from the sampled deep features to determine as the attention map of the shallow features, so as to obtain the deep features to be fused; A first execution submodule is used to multiply the deep features to be fused and the shallow features element by element to obtain the shallow features after suppressing background noise; The execution module includes: The third acquisition submodule is used to count the number of pixels whose characteristic response values are greater than 0 in the fused polyp characteristic image to obtain a first pixel value; A fourth acquisition submodule is used to count the number of pixels whose characteristic response values are less than 0 in the fused polyp characteristic image to obtain a second pixel value; The third execution submodule is used to normalize the first pixel value and the second pixel value, and divide the characteristic response value greater than 0 in the polyp characteristic image by the normalized first pixel value, and divide the characteristic response value less than 0 in the polyp characteristic image by the normalized second pixel value to obtain a corrected polyp characteristic image. 5. The segmentation device according to claim 4, characterized in that the processing module comprises: A first processing submodule is used to convert the colors of the polyp image X1 and the reference image X2 from the RGB color space to the LAB color space to obtain color values L1 and L2 of the polyp image X1 and the reference image X2 in the LAB color space; The second processing submodule is used to calculate the mean of the channels of the polyp image X1 in the LAB color space and standard deviation And the mean value of the channel of the reference image X2 in the LAB color space/> and standard deviation/> ; The third processing submodule is used to obtain the color value of the polyp image Y1 in the RGB color space and the color value of the reference image Y2 in the RGB color space by using a preset color conversion formula. 6. The segmentation device according to claim 4, characterized in that the processing module comprises: The second acquisition submodule is used to extract the first feature, the second feature and the third feature of the last three scales when the shallow features after suppressing the background noise are processed by the convolutional neural network; a fifth processing submodule, configured to fuse the first feature and the second feature to obtain a first fused feature; a sixth processing submodule, configured to fuse the second feature and the third feature to obtain a second fused feature; The second execution submodule is used to splice the first fusion feature and the second fusion feature according to channels to obtain a final fusion feature.