An improved U 2 -Net-based salient Mongolian pattern detection algorithm

Abstract

The invention discloses an improved U 2 The significant Mongolian pattern detection algorithm based on ‑Net includes: obtaining an image to be detected; inputting the image to be detected into a significant Mongolian pattern detection model, and outputting a significant Mongolian pattern detection result; wherein the significant Mongolian pattern detection model is based on an improved U 2 The deep learning model of the ‑Net network includes an encoder, a decoder, a feature supplement extraction module, a lightweight combined attention module and a feature map fusion module; the feature supplement extraction module adopts a small-size convolution kernel series connection method to increase the receptive field, reduce the computational cost, and make the predicted saliency map more complete; the lightweight combined attention module makes the boundary positioning of the saliency map more accurate without significantly increasing the computational cost, while suppressing redundant background noise, solving the technical problems of incomplete prediction of saliency Mongolian patterns, inaccurate edges and low detection accuracy.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and more particularly to an improved U 2 -Net-based salient Mongolian pattern detection algorithm. Background Art

[0002] Mongolian patterns are an important part of the traditional culture of the nation. They are formed by the cultural accumulation of the Mongolian people on horseback for thousands of years. They are the externalized expression of Mongolian thoughts and emotions and the unique visual symbols of the Mongolian people. They are one of the most representative intangible cultural heritages in the culture loved by the Mongolians. Therefore, in order to facilitate the secondary design and development of the inheritors of intangible cultural heritage, the significant Mongolian patterns can be extracted through significance testing.

[0003] The task of saliency detection is to use image processing technology and computer vision algorithms to separate the most eye-catching or important areas in the image from the background of the image. It mainly analyzes the salient areas of the image through feature extraction and calculation of feature contrast to obtain the detection results.

[0004] In the prior art, the research on the detection of single body regions in sample image saliency detection has been relatively mature. However, since Mongolian patterns are more complex, compared with conventional salient target detection tasks, higher requirements are placed on the integrity and edge details of image extraction results.

[0005] U 2 -Net is a simple and powerful deep network architecture proposed for the task of salient object detection, and a new module called RSU (Residual U-block) is proposed. 2 -Net network will frequently downsample (i.e. reduce the image) the feature map during the feature extraction stage of the whole network and each RSU. Although this operation will reduce memory and computing resources, it is also very likely to lose some important details of the image in the early stage of feature extraction, which will lead to incomplete prediction of significant Mongolian patterns and inaccurate edges. This will reduce the accuracy of significant Mongolian pattern detection to a certain extent, and cause certain negative impacts on the secondary design and development of Mongolian patterns in the later stage.

[0006] Therefore, an improved U 2 -Net architecture-based salient Mongolian pattern detection algorithm is very necessary. Summary of the invention

[0007] The present application embodiment provides an improved U 2-Net's salient Mongolian pattern detection algorithm is used to solve the technical problems of incomplete prediction of salient Mongolian patterns, inaccurate edges, and low detection accuracy.

[0008] In view of this, the present application provides an improved 2 -Net’s salient Mongolian pattern detection algorithm, including:

[0009] Acquire the image to be detected;

[0010] The image to be detected is input into the significant Mongolian pattern detection model, and the significant Mongolian pattern detection result is output; the significant Mongolian pattern detection model is based on the improved U 2 -Net network’s deep learning model;

[0011] Based on the improved U 2 -Net network's deep learning model includes encoder, decoder, feature supplement extraction module, lightweight combined attention module and feature map fusion module;

[0012] The encoder is a six-level encoder, and the encoder is a structure composed of multiple RSUs;

[0013] The feature supplement extraction module is located between two adjacent RSUs and is used to supplement the important features lost during the sampling process. The feature supplement extraction module can be divided into three groups, namely the first feature supplement extraction module, the second feature supplement extraction module and the third feature supplement extraction module. The inputs of the three groups of feature supplement extraction modules are residually connected with the outputs of the upper RSU. The outputs of the three groups of feature supplement extraction modules are fused with the outputs of the upper RSU, and the fused features are input into the lightweight combined attention module.

[0014] The lightweight combined attention module is used to enhance the ability to accurately locate salient patterns, including a channel attention module and a spatial attention module. The channel attention module can be multiplied with the output result of the spatial attention module to re-weight the features output by the upper-layer RSU structure.

[0015] Optionally, the three groups of feature supplement extraction modules have a total of five sub-modules, wherein the first feature supplement extraction module includes two sub-modules, the second feature supplement extraction module includes one sub-module, and the third feature supplement extraction module includes two sub-modules.

[0016] Optionally, the two submodules of the first complementary feature extraction module have the same structure, and the two submodules of the third complementary feature extraction module have the same structure.

[0017] Optionally, the submodule of the first feature supplementation extraction module includes a two-dimensional convolutional layer, a normalization layer and a ReLU layer connected in sequence.

[0018] Optionally, the submodule of the second feature supplementation extraction module includes a two-dimensional convolution layer, a normalization layer and a Relu layer, and the number of the two-dimensional convolution layer, the normalization layer and the Relu layer is two.

[0019] Optionally, the submodule of the third feature supplementation extraction module includes a two-dimensional convolution layer, a normalization layer and a Relu layer, and the number of the two-dimensional convolution layer, the normalization layer and the Relu layer is three.

[0020] Optionally, the convolution kernel size in the 2D convolution layer is 3x3.

[0021] Optionally, the channel attention module includes a global average pooling layer, a one-dimensional convolution layer and an activation layer. The global average pooling layer performs a global average pooling operation on the fused features to obtain a feature strip. The feature strip enters the one-dimensional convolution layer for feature extraction and then enters the activation layer. The activation layer obtains the weight of each feature point through a Sigmoid function, and finally multiplies it with the fused feature map pixels to obtain a feature map weighted by the channel information.

[0022] Optionally, the spatial attention module includes a global maximum pooling layer, a global average pooling layer, a feature layer, a two-dimensional convolution layer and an activation layer, performs global average pooling and global maximum pooling operations on the feature map weighted by the channel information, concatenates the results to obtain intermediate features, performs convolution operations on the intermediate features and obtains corresponding weights through the Sigmoid function, and finally multiplies the pixels of the feature map weighted by the channel information to obtain the significant Mongolian pattern detection result.

[0023] Optionally, the convolution kernel size of the 2D convolution layer in the spatial attention module is 7x7.

[0024] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0025] The present invention is based on U 2 -Net network’s deep learning model was improved to build a significant Mongolian pattern detection model. The significant Mongolian pattern detection model was used in U 2-Net network, a feature supplement extraction module is added between the RSUs. The small-size convolution kernel is connected in series in the feature supplement extraction module, which can not only increase the receptive field, but also reduce the computational cost in the downsampling process, effectively shorten the amount of calculation and improve the feature extraction capability, and the predicted saliency map is more complete; a lightweight combined attention module consisting of a channel attention module and a spatial attention module is added, which can effectively enhance the response of salient targets in the features without significantly increasing the computational cost, and make the prediction of the boundaries of salient areas more accurate while suppressing redundant background noise, and finally obtain a salient Mongolian pattern with precise boundaries; the technical problems of incomplete prediction of salient Mongolian patterns, inaccurate edges and low detection accuracy are solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 The improved U provided in the embodiment of the present invention 2 -Schematic diagram of the process of detecting significant Mongolian patterns based on Net;

[0027] Figure 2 The improved U provided in the embodiment of the present invention 2 -Net network structure diagram;

[0028] Figure 3 It is a structural schematic diagram of a feature supplementation extraction module provided in an embodiment of the present invention;

[0029] Figure 4 A schematic diagram of the structure of a lightweight combined attention module provided in an embodiment of the present invention;

[0030] Figure 5 The improved U provided in the embodiment of the present invention 2 Comparison of the detection results of -Net's salient Mongolian pattern detection method and other detection methods. DETAILED DESCRIPTION

[0031] In order to enable those skilled in the art to better understand the scheme of the present application, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0032] For easier understanding, see Figure 1 and Figure 2 (in Figure 2 Where FSE is a feature supplement extraction module, LCA is a lightweight combined attention module, En is an encoder, and De is a decoder), the present application provides an improved U 2 An embodiment of a salient Mongolian pattern detection algorithm of -Net, comprising:

[0033] S1: Get the image to be detected;

[0034] S2: Input the image to be detected into the significant Mongolian pattern detection model, and output the significant Mongolian pattern detection result; the significant Mongolian pattern detection model is based on the improved U 2 -Net network’s deep learning model;

[0035] Based on the improved U 2 -Net network's deep learning model includes encoder, decoder, feature supplement extraction module, lightweight combined attention module and feature map fusion module;

[0036] The encoder is a six-level encoder, and the encoder is a structure composed of multiple RSUs;

[0037] The first four encoding stages of the encoder (En_1, En_2, En_3, and En_4) use the RSU structure of RSU-7, RSU-6, RSU-5, and RSU-4, respectively (where "7", "6", "5", and "4" represent the height of RSU). For feature maps with larger height and width, the upper layer uses a larger height to capture information of a larger scale. In the fifth and sixth encoding stages (En_5 and En_6), the resolution of the feature map is relatively low, and RSU-4F is used, where F indicates that RSU is an extended version, and the expansion convolution is used to replace the merging and upsampling operations, so that all intermediate feature maps of RSU-4F have the same resolution as their input feature maps, and downsampling the feature map leads to the loss of useful context.

[0038] The feature supplement extraction module is located between two adjacent RSUs and is used to supplement the important features lost during the sampling process. The feature supplement extraction module can be divided into three groups, namely the first feature supplement extraction module, the second feature supplement extraction module and the third feature supplement extraction module. The inputs of the three groups of feature supplement extraction modules are residually connected with the outputs of the upper RSU. The outputs of the three groups of feature supplement extraction modules are fused with the outputs of the RSU and enter the lightweight combined attention module.

[0039] The three groups of feature supplementary extraction modules have a total of five submodules, among which the first group of feature supplementary extraction modules includes two submodules (FSE_1-1 and FSE_1-2), the second group of feature supplementary extraction modules includes one submodule (FSE_2-1), and the third group of feature supplementary extraction modules includes two submodules (FSE_3-1 and FSE_3-2); FSE_1-1 is located between En_1 and En_2, FSE_1-2 is located between En_2 and En_3, FSE_2-1 is located between En_3 and En_4, FSE_3-1 is located between En_4 and En_5, and FSE_3-2 is located between En_5 and En_6;

[0040] See also Figure 3 The five sub-modules of the three groups of feature supplementation extraction modules are respectively composed of different numbers of convolution kernels connected in series. The FSE_1-1 and FSE_1-2 in the first feature supplementation extraction module have the same structure, both of which include a two-dimensional convolution layer, a normalization layer and a Relu layer connected in sequence. The FSE_2-1 of the second feature supplementation extraction module includes a two-dimensional convolution layer, a normalization layer and a Relu layer, wherein the number of the two-dimensional convolution layer, the normalization layer and the Relu layer is two. The FSE_3-1 and FSE_3-2 in the third feature supplementation extraction module have the same structure, both of which include a two-dimensional convolution layer, a normalization layer and a Relu layer, wherein the number of the two-dimensional convolution layer, the normalization layer and the Relu layer is three. The convolution kernel size of the convolution layer of the five sub-modules in the three groups of feature supplementation extraction modules is 3x3. The inputs of the five sub-modules of the three groups of feature supplementation extraction modules are connected to the outputs of the upper-layer RSU through residual connections.

[0041] When the feature map enters the first feature supplement extraction module, a convolution kernel of size 3x3 is first used to extract features from the input feature map. At this time, the number of channels will be adjusted to be the same as the feature map output by the adjacent RSU, and then the feature is normalized and the Relu activation function is used to set all negative numbers to 0. The main operations of the second and third feature supplement extraction modules are exactly the same as those of the first feature supplement extraction module, except that a series of operations of the first feature supplement extraction module are repeated 2 and 3 times to realize the convolution kernel series operation. As the network deepens and continues to downsample, the size of the feature map decreases. Therefore, the number of series convolution kernels is progressively increased to obtain a larger receptive field, thereby extracting high-level structural features, which can effectively avoid the loss of important features during the downsampling process. Finally, the outputs of the five submodules in the three groups of feature supplement extraction modules are fused with the output of the RSU and input into the lightweight combined attention module.

[0042] In the feature supplementation extraction module, the serial convolution operation of small-size convolution kernels can increase the receptive field while reducing the computational cost. For example, using a 3x3 convolution kernel for two consecutive convolutions can achieve the ability to extract feature maps by using a 5x5 convolution kernel for one convolution (i.e., the second feature supplementation extraction module convolves twice). Similarly, using a 3x3 convolution kernel for three consecutive convolutions can achieve the ability to extract feature maps by using a 7x7 convolution kernel for one convolution (i.e., the third feature supplementation extraction module convolves three times). However, replacing one 5x5 convolution with two 3x3 cascaded convolutions can reduce 5*5*L-2*3*3*L=7*L calculations, reducing the amount of calculation by 28%. Replacing one 7x7 convolution with three 3x3 cascaded convolutions can reduce 7*7*L-3*3*3*L=22*L calculations, reducing the amount of calculation by 45%, effectively shortening the amount of calculation and improving the feature extraction capability. Adding normalization layers and Relu activation functions can accelerate the convergence of the network model and avoid network generalization.

[0043] In the existing U 2 -Net network, continuous downsampling by multiple RSUs may easily cause the loss of important features. At the same time, the lost important features cannot be supplemented in the subsequent upsampling process, so that the feature map obtained after upsampling is still incomplete; the feature supplement extraction module is set between adjacent RSUs. After downsampling, the feature supplement extraction module can further extract features, and then fuse them with the output features of RSU to avoid the loss of important features; and the feature supplement extraction module can not only supplement the lost features but also fuse the features extracted from different receptive fields, so that the saliency map predicted by this network model is more complete.

[0044] See also Figure 4 , a lightweight combined attention module is used to enhance the ability of accurate positioning of salient patterns, including a channel attention module and a spatial attention module, wherein the channel attention module includes a global average pooling layer, a one-dimensional convolution layer and an activation layer, and the spatial attention module includes a global maximum pooling layer, a global average pooling layer, a feature layer, a two-dimensional convolution layer and an activation layer. The channel attention module can be multiplied with the output result of the spatial attention module to re-weight the features output by the upper RSU structure, thereby optimizing the prediction details of the salient map while suppressing noise;

[0045] The feature map is located at U 2 After the RSU of the -Net encoder structure is fused with the feature supplement extraction module, it will be sent to the adjacent lightweight combined attention module. The lightweight combined attention module will re-weight the output results of En_i, thereby optimizing the prediction details of the saliency map while suppressing noise.

[0046] The channel attention module in the lightweight combined attention module first performs a global average pooling operation on the input features in the global average pooling layer to obtain a feature strip of size 1*1*Cout(i), where Cout(i) represents the number of channels after the output fusion of the i-th RSU module and the feature supplementary extraction module; the feature strip enters the one-dimensional convolution layer for feature extraction. The convolution layer uses a one-dimensional convolution operation, and there is a certain mapping relationship between the convolution kernel size K and Cout(i):

[0047]

[0048] The default values ​​of r and b are 2 and 1, and the feature length after feature extraction is still Cout(i). Then, the activation layer uses the Sigmoid function to obtain the weight of each feature point on the extracted features, and finally multiplies it with the input feature map pixel of the channel attention module to obtain the feature map weighted by the channel information.

[0049] Different from the traditional multi-layer fully connected layer feature extraction method, when the traditional multi-layer fully connected layer performs feature extraction, for complex pattern detection, it is necessary to increase the number of neurons to improve the learning ability of the model, which is likely to cause overfitting, and also lead to increased computing time and reduced efficiency; the present invention adopts the convolution operation of the channel attention module to extract features from feature strips. Since the convolution operation has good cross-channel information acquisition ability, it not only has a global receptive field, but also can extract rich high-level features.

[0050] The feature map weighted by channel information is sent to the spatial attention module. First, global average pooling and global maximum pooling operations are performed on the channels of the feature map weighted by channel information. Then, the results are concatenated to generate an intermediate feature of size H(i)*W(i)*2, where H(i) and W(i) represent the height and width of the fused output of the i-th RSU and the feature supplement extraction module. The intermediate feature is convolved with a convolution kernel of size 7 to adjust its channel number to 1. The corresponding weight is obtained in the activation layer through the sigmoid function. Finally, the weight output by the spatial attention module is multiplied by the feature map pixel output by the channel attention module to obtain the output feature.

[0051] Without significantly increasing the computational cost, the lightweight combined attention module used in the model can more effectively enhance the response of salient targets in the features, thereby suppressing redundant background noise while making the prediction of the boundaries of salient regions more accurate, and ultimately obtaining a salient Mongolian pattern with precise boundaries.

[0052] The decoding stage has a similar structure to the encoding stage, including five decoding stages (De_1, De_2, De_3, De_4, De_5). In De_5, an extended version of RSU-4F is used, which is similar to that used in the encoding stage En_5, En_6. Each decoder stage takes the concatenation of the feature map fused by the lightweight attention module and the symmetric encoder stage and the upsampled feature map from the previous stage as input, and its output is fed into the feature map fusion module.

[0053] The feature map fusion module is used to horizontally fuse the weighted features of the lightweight combined attention module with the features of each part of the decoder, and then generate six output significant probability maps through 3x3 convolution and Sigmoid function. Then, the logical map of the output feature map (convolution output, before Sigmoid function) is upsampled to the same size as the input image, and fused through the cascade operation concat, and then through a 1x1 convolution layer and a Sigmoid function to generate the final significant Mongolian pattern probability map, and output the significant Mongolian pattern detection result.

[0054] See also Figure 5 , the present invention will be based on the improved U 2 -Net salient Mongolian pattern detection model is compared with the detection results obtained by the five most advanced open source detection methods in the prior art. It can be seen that the detection effect and the true value map of the model of the present invention are highly consistent when the salient pattern is complex, the background is complex, and the similarity between the salient area and the background is high. In addition, compared with other methods, U 2 -Net has almost no defects when the edge of the salient pattern is complex, and can detect a more complete salient pattern outline; in the case of a complex background, it can eliminate background interference and effectively detect the salient pattern; when the similarity between the salient area and the background is high, it can also accurately separate the background information and find the salient pattern well.

[0055] The detection results obtained by other methods show that when the edges of the significant Mongolian patterns are complex, defects are prone to occur, making the contour of the significant pattern unclear and the Mongolian pattern extraction result incomplete; when the significant area is highly similar to the background, the interference information of the background is easily detected together, and the significant pattern obtained is not accurate enough.

[0056] It should be noted that the original U 2-Net network uses multiple U-Net-like structures (RSU) in the encoder structure of the feature extraction stage. This structure will alternately perform multiple two-dimensional convolutions with a convolution kernel size of 3x3, upsampling, and downsampling operations on the original input features. For example, if the size of the original image is 512x512, it will become five different sizes of features: 256x256, 128x128, 64x64, 32x32, and 16x16. Then the 16x16 feature map is upsampled to obtain a 32x32 feature map. This 32x32 feature map is spliced ​​with the previous 32x32 feature map by channel damage, and then the spliced ​​feature map is convolved and upsampled to obtain a 64x64 feature map, which is then spliced ​​with the previous 64x64 feature, convolved, and upsampled. After 5 upsamplings, a 512x512 prediction result with the same size as the input image can be obtained. Although this structure can increase the receptive field of feature extraction as the network structure deepens, the frequent downsampling of feature maps will also lead to the loss of original important features. After subsequent upsampling and feature concatenation, the lost important features still cannot be effectively supplemented, making the final prediction result inaccurate. Therefore, this application uses the original U 2 -Net network structure, a feature supplement extraction module and a lightweight combined attention module are added to obtain more accurate Mongolian pattern detection results.

[0057] The significant Mongolian pattern detection model provided in the embodiment of the present application is based on U 2 -Net network's deep learning model is improved, and the significant Mongolian pattern detection model is used in U 2 -Net network, a feature supplement extraction module is added between the RSUs. The small-size convolution kernel is connected in series in the feature supplement extraction module, which can not only increase the receptive field, but also reduce the computational cost in the downsampling process, effectively shorten the amount of calculation and improve the feature extraction capability, and the predicted salient map is more complete; a lightweight combined attention module consisting of a channel attention module and a spatial attention module is added, which can effectively enhance the response of salient targets in the features, thereby suppressing redundant background noise while making the prediction of the boundaries of salient areas more accurate, and finally obtaining salient Mongolian patterns with precise boundaries; the technical problems of incomplete prediction of salient Mongolian patterns, inaccurate edges and low detection accuracy are solved.

[0058] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An improved U 2 -Net's salient Mongolian pattern detection algorithm, characterized by: include: Acquire the image to be detected; Inputting the image to be detected into a significant Mongolian pattern detection model, and outputting a significant Mongolian pattern detection result; The significant Mongolian pattern detection model is based on the improved U 2 -Net network’s deep learning model; The improved U 2 -Net network's deep learning model includes encoder, decoder, feature supplement extraction module, lightweight combined attention module and feature map fusion module; The encoder is a six-level encoder, and the encoder is a structure composed of multiple RSUs; the first four encoding stages of the encoder use RSU structures of RSU-7, RSU-6, RSU-5, and RSU-4 respectively; wherein the number 7 in RSU-7, the number 6 in RSU-6, the number 5 in RSU-5, and the number 4 in RSU-4 represent the height of the RSU; for feature maps with large height and width, the upper RSU uses a large height; the fifth and sixth encoding stages of the encoder use the RSU-4F structure, and the F in RSU-4F indicates that the RSU is an extended version, and the extended convolution is used instead of the merging and upsampling operations; The feature supplement extraction module is located between two adjacent RSUs and is used to supplement and extract important features lost during the sampling process. The feature supplement extraction module can be divided into three groups, namely, a first feature supplement extraction module, a second feature supplement extraction module and a third feature supplement extraction module. The inputs of the three groups of feature supplement extraction modules are residually connected with the outputs of the upper RSU, and the outputs of the three groups of feature supplement extraction modules are fused with the outputs of the upper RSU, and the fused features are input into the lightweight combined attention module; The operations of the second feature supplementation extraction module and the third feature supplementation extraction module are the same as the operations of the first feature supplementation extraction module. By repeating the operations of the first feature supplementation extraction module multiple times, a series connection of small-size convolution kernel operations is realized, and the series connection of small-size convolution kernel operations can reduce the calculation cost; The lightweight combined attention module is used to enhance the ability of accurately locating significant patterns, and includes a channel attention module and a spatial attention module. The channel attention module can be multiplied with the output result of the spatial attention module to re-weight the features output by the upper-layer RSU structure.

2. An improved U according to claim 1 2 -Net's salient Mongolian pattern detection algorithm, characterized by: The three groups of feature supplement extraction modules have a total of five sub-modules, wherein the first feature supplement extraction module includes two sub-modules, the second feature supplement extraction module includes one sub-module, and the third feature supplement extraction module includes two sub-modules.

3. An improved U according to claim 2 2 -Net's salient Mongolian pattern detection algorithm, characterized by: The two submodules of the first feature supplementation extraction module have the same structure, and the two submodules of the third feature supplementation extraction module have the same structure.

4. An improved U according to claim 3 2 -Net's salient Mongolian pattern detection algorithm, characterized by: The submodule of the first feature supplementation extraction module includes a two-dimensional convolutional layer, a normalization layer and a ReLU layer connected in sequence.

5. An improved U according to claim 2 2 -Net's salient Mongolian pattern detection algorithm, characterized by: The submodules of the second feature supplementation extraction module include a two-dimensional convolution layer, a normalization layer and a Relu layer, and the number of the two-dimensional convolution layer, the normalization layer and the Relu layer is two.

6. An improved U according to claim 3 2 -Net's salient Mongolian pattern detection algorithm, characterized by: The submodules of the third feature supplementation extraction module include a two-dimensional convolution layer, a normalization layer and a Relu layer, and the number of the two-dimensional convolution layer, the normalization layer and the Relu layer is three.

7. An improved U according to any one of claims 4 to 6. 2 -Net's salient Mongolian pattern detection algorithm, characterized by: The convolution kernel size in the two-dimensional convolution layer is 3x3.

8. An improved U according to claim 1 2 -Net's salient Mongolian pattern detection algorithm, characterized by: The channel attention module includes a global average pooling layer, a one-dimensional convolution layer and an activation layer. The global average pooling layer performs a global average pooling operation on the fused features to obtain a feature strip. The feature strip enters the one-dimensional convolution layer for feature extraction and then enters the activation layer. The activation layer obtains the weight of each feature point through a Sigmoid function, and finally multiplies it with the fused feature map pixels to obtain a feature map weighted by channel information.

9. An improved U according to claim 8. 2 -Net's salient Mongolian pattern detection algorithm, characterized by: The spatial attention module includes a global maximum pooling layer, a global average pooling layer, a feature layer, a two-dimensional convolution layer and an activation layer. The feature map weighted by the channel information is subjected to global average pooling and global maximum pooling operations, the obtained results are concatenated to obtain intermediate features, the intermediate features are convolved and the corresponding weights are obtained through the Sigmoid function, and finally multiplied with the feature map pixels weighted by the channel information to obtain the significant Mongolian pattern detection result.

10. An improved U according to claim 9. 2 -Net's salient Mongolian pattern detection algorithm, characterized by: The convolution kernel size of the two-dimensional convolution layer in the spatial attention module is 7x7.

Images