High-resolution image target identification and segmentation method

Abstract

The invention discloses a high-resolution image target identification and segmentation method. The method comprises the following steps: performing slice division on a preset original image according to the same size to obtain a plurality of image slices; performing feature mapping on the plurality of image slices through a preset full-connection network to obtain a plurality of initial vectors; performing feature extraction on the initial vector to obtain a fused feature vector; performing multi-level feature extraction on the fusion feature vector to obtain a high-dimensional fusion feature vector; and performing up-sampling on the high-dimensional fusion feature vector to obtain a semantic segmentation result graph. According to the method, the problems that in the prior art, the network cannot fully extract the high-dimensional features of the high-resolution image and the spatial modeling capability is not high are solved, and the recognition precision of the target in the high-resolution image is further improved.

Description

Technical Field

[0001] The invention belongs to the technical field of image target recognition and segmentation, and in particular relates to a high-resolution image target recognition and segmentation method. Background Art

[0002] Image object recognition and segmentation is an important research direction in the field of computer vision. It involves identifying and locating objects of a specific category in an image or video. This technology has a wide range of applications, including military, automation, security monitoring, and autonomous driving. In the early days, object recognition relied on manually designed feature extractors, such as Hear features and HOG features, combined with machine learning classifiers such as SVM and AdaBoost to identify and classify objects. With the successful application of neural networks in image classification tasks, the field of object recognition has ushered in a revolutionary change. Neural networks, especially convolutional neural networks, can learn complex feature representations from data, greatly improving the accuracy and efficiency of object recognition. Object recognition in high-resolution images (images with a resolution of 2160×1080 and above) aims to identify and locate subtle features of specific targets (such as target substructures, edge features, etc.). By segmenting the image into multiple regions, each region represents a specific category, the classification of each pixel in the image can be achieved. This technology can provide a deeper understanding of scene information and obtain more useful information.

[0003] Involving patent [Deep learning image segmentation method and device based on sparse features, 201910730849.7], the semantic information segmentation result corresponding to the original image is obtained by combining the coordinate information of the target point of the target feature image segmentation. For high-resolution images, the target feature image generated by this method does not extract sufficient features, and the scale transformation will cause the loss of the edge features of the target object, which will eventually affect the recognition accuracy. Involving patent [Image segmentation and model training method, device and equipment, 202211528536.1], this method generates segmentation results through an image encoder and a classification-guided decoder. This method does not fully extract image features and does not consider the spatial relationship between features at different scales. Involving patent [An image segmentation network and image segmentation method, 202111606915.3], this method realizes feature extraction and fusion through a feature pyramid structure, and uses an attention module to extract the features of the target object. This method does not pay attention to image features at different scales and different spatial orientations, the feature pyramid has insufficient spatial modeling capabilities for images, the extracted features are not rich enough, and the recognition accuracy for high-resolution images is not high.

[0004] In summary, existing technologies can effectively identify targets. However, for complex target substructures or target edge features in high-resolution images, the recognition accuracy needs to be improved due to insufficient ability to extract spatial features of images, especially the inadequate establishment of the correlation between features at different scales. Summary of the invention

[0005] The technical problem solved by the present invention is: to overcome the shortcomings of the prior art, to provide a method for high-resolution image target recognition and segmentation, to solve the problems in the prior art of insufficient network extraction of high-dimensional features of high-resolution images and weak spatial modeling ability, and to further improve the recognition accuracy of targets in high-resolution images.

[0006] The object of the present invention is achieved through the following technical scheme: a high-resolution image target recognition and segmentation method, comprising: slicing a preset original image according to the same size to obtain multiple image slices; performing feature mapping on the multiple image slices through a preset fully connected network to obtain multiple initial vectors; performing feature extraction on the initial vectors to obtain a fused feature vector; performing multi-level feature extraction on the fused feature vector to obtain a high-dimensional fused feature vector; and upsampling the high-dimensional fused feature vector to obtain a semantic segmentation result map.

[0007] In the above-mentioned high-resolution image target recognition and segmentation method, high-dimensional feature extraction of the initial vector is performed to obtain a fused feature vector, including: performing a normalization layer calculation on the initial vector to obtain a normalized vector; establishing a window according to the normalized vector; sliding the window on the normalized vector and performing self-attention calculation to obtain a sliding window self-attention vector; performing a mask matrix calculation on the sliding window self-attention vector to obtain mask vectors in three directions; and performing attention feature fusion on the mask vectors in the three directions to obtain a fused feature vector.

[0008] In the above high-resolution image target recognition and segmentation method, the normalized vector includes a mean and a variance.

[0009] In the above high-resolution image target recognition and segmentation method, the normalized vector is obtained by the following formula:

[0010]

[0011]

[0012] Among them, μ i is the mean of the i-th initial vector, σ i is the variance of the ith initial vector, W is the number of initial vectors, x i is the i-th initial vector, ε is the noise, and i is the number of the initial vector.

[0013] In the above high-resolution image target recognition and segmentation method, the sliding window self-attention vector is obtained by the following formula:

[0014]

[0015] Among them, z l+1 is the sliding window self-attention vector, MLP() represents the multi-layer perceptron, WMSA() and SWMSA() represent the window-based self-attention calculation function and the sliding window-based self-attention calculation function respectively, LN() represents the normalization layer calculation function, is the first intermediate vector, z l is the second intermediate vector, is the third intermediate vector, and l is the window number.

[0016] In the above high-resolution image target recognition and segmentation method, the three directions are the upper left, lower left, and upper right directions of the window sliding.

[0017] In the above high-resolution image target recognition and segmentation method, the fused feature vector is:

[0018] Among them, Q is the mask vector in the lower left direction, K is the mask vector in the upper left direction, V is the mask vector in the upper right direction, and d is the scaling factor.

[0019] A high-resolution image target recognition and segmentation system comprises: a first module, used for slicing a preset original image into slices of the same size to obtain a plurality of image slices; performing feature mapping on the plurality of image slices through a preset fully connected network to obtain a plurality of initial vectors; a second module, used for performing feature extraction on the initial vectors to obtain a fused feature vector; a third module, used for performing multi-level feature extraction on the fused feature vector to obtain a high-dimensional fused feature vector; and a fourth module, used for upsampling the high-dimensional fused feature vector to obtain a semantic segmentation result map.

[0020] In the above-mentioned high-resolution image target recognition and segmentation system, high-dimensional feature extraction of the initial vector is performed to obtain a fused feature vector, including: performing normalization layer calculation on the initial vector to obtain a normalized vector; establishing a window based on the normalized vector; sliding the window on the normalized vector and performing self-attention calculation to obtain a sliding window self-attention vector; performing mask matrix calculation on the sliding window self-attention vector to obtain mask vectors in three directions; and performing attention feature fusion on the mask vectors in the three directions to obtain a fused feature vector.

[0021] In the above high-resolution image target recognition and segmentation system, the normalized vector includes a mean and a variance.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] (1) The present invention performs two operations, slicing and word embedding layer assignment, on the original high-resolution image. Slicing divides the high-resolution image into small-scale tensors, which is convenient for network processing and calculation;

[0024] (2) Based on the modeling of windows and window sliding of images, the present invention expands the network's receptive field and enhances the network's ability to extract features from images based on the multi-directional movement of windows in the image matrix;

[0025] (3) The present invention uses the attention mechanism to fuse the features extracted by sliding windows in different directions. Through the attention mechanism, the network can learn the association information between windows in different directions on the image, which helps to improve the network's recognition accuracy for specific target substructures;

[0026] (4) The mask matrix of the present invention can separate the associated area and the non-associated area after the window slides by setting the maximum and minimum values, so that the features extracted by the network after the window slides are also relevant, avoiding the gradient drift phenomenon during the network training process. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0028] Figure 1 is a flow chart of a high-resolution image target recognition and segmentation method provided by an embodiment of the present invention;

[0029] Figure 2 It is a schematic diagram of a multi-directional sliding window structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0030] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to be able to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features described in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0031] Figure 1This embodiment provides a method for high-resolution image target recognition and segmentation, combined with Figure 1 , the method comprising:

[0032] The preset original image is sliced ​​and divided into multiple image slices according to the same size; the multiple image slices are feature mapped through a preset fully connected network to obtain multiple initial vectors;

[0033] Perform feature extraction on the initial vector to obtain a fused feature vector;

[0034] Perform multi-level feature extraction on the fused feature vector to obtain a high-dimensional fused feature vector;

[0035] The high-dimensional fused feature vector is upsampled to obtain the semantic segmentation result map.

[0036] Performing high-dimensional feature extraction on the initial vector to obtain a fused feature vector includes: performing normalization layer calculation on the initial vector to obtain a normalized vector; establishing a window according to the normalized vector; sliding the window on the normalized vector and performing self-attention calculation to obtain a sliding window self-attention vector; performing mask matrix calculation on the sliding window self-attention vector to obtain mask vectors in three directions; and performing attention feature fusion on the mask vectors in the three directions to obtain a fused feature vector.

[0037] The normalized vector includes the mean and variance. The normalized vector is obtained by the following formula:

[0038]

[0039] Among them, μ i is the mean of the i-th initial vector, σ i is the variance of the ith initial vector, W is the number of initial vectors, x i is the i-th initial vector, ε is the noise, and i is the number of the initial vector.

[0040] The sliding window self-attention vector is obtained by the following formula:

[0041]

[0042] Among them, z l+1 is the sliding window self-attention vector, MLP() represents the multi-layer perceptron, WMSA() and SWMSA() represent the window-based self-attention calculation function and the sliding window-based self-attention calculation function respectively, LN() represents the normalization layer calculation function, is the first intermediate vector, z l is the second intermediate vector, is the third intermediate vector, and l is the window number.

[0043] The three directions are upper left, lower left, and upper right directions for window sliding.

[0044] The fused feature vector is: Among them, Q is the mask vector in the lower left direction, K is the mask vector in the upper left direction, V is the mask vector in the upper right direction, and d is the scaling factor.

[0045] Specifically, the method comprises the following steps:

[0046] Step 1: Slice the original image into slices of the same size to obtain multiple image slices; perform feature mapping on each image slice through a preset fully connected network to obtain each initial vector.

[0047] Assume that the size of the original image is n×m, the number of slices required is x, and the original image is divided into slices of the same size according to the size of (n / x)×(m / x). Next, the values ​​of the first two dimensions of each slice are arranged in the spatial order of the third dimension, and the result is feature mapped through a fully connected network to obtain a feature result containing word embedding information. Finally, the mapped features are passed as input to step 2.

[0048] Step 2: Extract fusion feature vector

[0049] The initial vector in step 1 is passed to the feature extraction stage for high-dimensional feature extraction. Figure 2 As shown in the figure, the feature extraction stage consists of normalization layer calculation, sliding window self-attention calculation, mask matrix calculation, and channel attention feature fusion.

[0050] Step 2.1: Perform a normalization layer calculation on the initial vector in step 1 to obtain a normalized vector; wherein the normalized vector includes a mean and a variance.

[0051]

[0052] x i is the i-th initial vector, μ i is the mean of the i-th initial vector, σ i is the variance of the ith initial vector, W is the number of initial vectors, and ε is the noise.

[0053] The initial features are constrained to a standard normal distribution using mean variance and restored using trainable parameters.

[0054] After training the normalization layer, the multi-layer networks are decoupled to a certain extent, so that each layer of the network maintains a relatively stable distribution. At the same time, the final model is no longer sensitive to the parameters of the shallow network, and the back propagation process becomes more stable.

[0055] Step 2.2: Sliding window self-attention calculation

[0056] A window is established according to the normalized vector; wherein the size of the window is 2 dimensions in length and 2 dimensions in width; the window is slid on the normalized vector and self-attention calculation is performed to obtain the sliding window self-attention vector.

[0057] In the sliding window self-attention calculation stage, the multi-head self-attention within the window is calculated first, and then the multi-head self-attention of the sliding window is calculated. The two partition configurations of window and sliding window are used alternately in the continuous self-attention calculation modules. Assume that each window contains M×M image slices, and the window sliding steps are as follows: the first module uses the conventional window partitioning strategy, starting from the upper left corner, to evenly divide the 8×8 feature map into 2×2 windows with a window size of 4. Then, the next module shifts the window from the regularly divided window by ([M / 2], [M / 2]) positions and uses the same window configuration as the previous layer. The sliding distance of each sliding of the sliding window is equal (([M / 2], [M / 2]) positions), and a window-based self-attention calculation is performed after each change in direction to obtain the feature vector in that direction.

[0058] The calculation process of the continuous sliding window self-attention calculation module is as follows: after the correction of the normalization layer, the window self-attention calculation is performed first, and then the features are input into the multi-layer perceptron. Then, the output of the multi-layer perceptron is subjected to the normalization layer calculation again, and the output is subjected to the sliding window self-attention calculation. Finally, the result is output by the multi-layer perceptron.

[0059]

[0060]

[0061] Among them, z l and They represent the window-based output features and the multi-layer perceptron-based output features in module 1 respectively; WMSA and SWMSA represent the multi-head self-attention calculation module based on a conventional window and the multi-head self-attention calculation module based on a multi-directional sliding window respectively; LN represents the normalization layer calculation, and MLP represents the multi-layer perceptron.

[0062] The self-attention module is mapped to the original input size by the multi-layer perceptron and then output, with a GELU nonlinear activation function in the middle. There is a normalization layer before each multi-head self-attention module, each sliding window-based multi-head self-attention module, and each multi-layer perceptron module, and there is a residual connection after each module. The window is delineated to evenly divide the original image in a non-overlapping manner. The computational complexity of the multi-head self-attention model and based on an h×w size image slice is:

[0063] Ω(MSA)=4hwC 2 +2(hw) 2 C;

[0064] Ω(W-MSA)=4hwC 2 +2M 2 hwC.

[0065] Step 2.3: Mask matrix calculation

[0066] The mask matrix of the sliding window self-attention vector is calculated to obtain mask vectors in three directions. The three directions are the directions in which the window slides: upper left, lower left, and upper right.

[0067] Generate a mask matrix of the same dimension as the sliding window self-attention vector. In the same window, the relevant area before and after sliding is filled with 0, and the non-relevant area before and after sliding is filled with -100. Then add the mask matrix to the original feature matrix to get the mask vector.

[0068] The calculation method of the mask matrix is ​​as follows: Assume the window size is M×M (2 dimensions × 2 dimensions), and the window sliding distance is p. Observe from the perspective of the adjacent window, the window before sliding is 4 M×M windows. Taking the sliding at the lower right as an example, the number and size of windows after sliding become 4 p×p, 4 p×(Mp), and 1 M×M, and the size of the channel dimension does not change. The 1 p×p area in the upper left is recorded as area A, the p×(Mp) area on the left is recorded as area B, and the p×(Mp) area directly above is recorded as area C. Then slide area B to the right of the entire area, slide area C to the bottom of the entire area, and slide area A to the gap between the new positions of areas B and C to fill it. Re-dividing the window according to the size of M×M, the whole can be divided into 4 M×M windows. Calculating self-attention in these 4 windows means that each window not only contains the element association information in the window before sliding, but also contains the element association information in the adjacent window after sliding, which expands the receptive field of the neural network and effectively improves the ability to extract features. In addition, the original 9 windows are changed to 4 windows, which is consistent with the number of windows before sliding, and the amount of calculation does not increase, reducing the calculation cost. However, this will bring about the problem of position confusion. Calculating self-attention between irrelevant areas in the original image will greatly reduce the effectiveness and accuracy of the calculation results. In order to solve this problem, a mask matrix with the same dimension size as the sliding window self-attention vector is generated. In the same window, the relevant area is filled with 0, and the non-relevant area is filled with -100. Then, the mask matrix is ​​added to the original feature matrix to obtain the mask vector, so that the relevant area element values ​​of the feature matrix after the area slide are retained, and the element values ​​of the non-relevant area become very small values. The result of adding the mask matrix to the original feature matrix is ​​the mask vector.

[0069] Step 2.4: Channel attention feature fusion

[0070] The mask vectors in three directions are fused with attention features to obtain a fused feature vector.

[0071] Channel attention feature fusion is an effective feature fusion method that can help the model better utilize the information between different channels and improve the performance of the model. The idea of ​​channel attention feature fusion is to introduce an attention mechanism for each channel to calculate the contribution of the channel to the final output. Specifically, for each channel, a weight is calculated to indicate the importance of the channel to the final output. Then, the output of each channel is multiplied by the corresponding weight, and the outputs of all channels are added together to obtain the final feature representation. The advantage of channel attention feature fusion is that it can adaptively learn the importance of each channel, thereby better utilizing the information between different channels.

[0072] First, the feature matrices in multiple directions are cropped, and each feature matrix is ​​divided into three parts with equal distance in the third dimension, and the dimensions of each submatrix are the same. The feature matrix is ​​divided to reduce the dimensions of the two matrices to be multiplied and reduce the amount of calculation. Secondly, it is to improve the efficiency of feature extraction and reduce the bias error caused by a linear layer. Each divided sub-part does not directly participate in the calculation of channel attention. It will first be sampled in parallel using a linear layer with the same input and output dimensions to improve the network perception ability and reduce the impact of excessive or small element values ​​in special areas.

[0073] Let a submatrix of the lower left feature matrix be Q, and the corresponding upper left and upper right feature matrix submatrices be K and V. Then the channel attention is calculated as follows:

[0074]

[0075] Among them, d k is the scaling factor of the feature matrix, which is a hyperparameter. After the Softmax calculation, a regularization layer with a zero rate of 0.1 is used to correct the output result to reduce the overfitting phenomenon of this layer. After multiplication with the feature matrix submatrix on the upper right, the three fused feature submatrices are concatenated in the third dimension, and a regularization layer with a zero rate of 0.1 is continued to be used to improve the robustness of this layer. The three fused feature submatrices are concatenated to obtain the fused feature matrix.

[0076] Step 3: Perform multi-level feature extraction on the fused feature vector outputted in step 2 to obtain a high-dimensional fused feature vector.

[0077] Perform multi-level feature extraction on the fused feature matrix output by step 2 to achieve the ability to model image space at different scales. The entire network consists of 4 layers of the network structure described in step 2. Based on the fused feature matrix output by step 2, the feature matrix is ​​halved in the first and second dimensions, and correspondingly, it is doubled in the third dimension. After each fusion feature extraction stage, the matrix dimension of the fused feature matrix will undergo the above changes. From the perspective of spatial structure, the cross-section of the feature matrix is ​​reduced and the length is increased. In order to enable the network to extract features at different spatial scales, after each fusion feature extraction stage, the adjacent 4 feature slices are merged so that the deep network can extract features in a larger scale space.

[0078] Assume that the original image size is H×W×3. After the first layer of fusion feature extraction, the fusion feature matrix size becomes After the second-layer fusion feature extraction stage, the fusion feature matrix size becomes After the third layer of fusion feature extraction stage, the size of the fusion feature matrix becomes After the fourth fusion feature extraction stage, the size of the fusion feature matrix becomes

[0079] Step 4: Upsample the high-dimensional fused feature vector to obtain the semantic segmentation result map. Upsampling is the deconvolution operation.

[0080] The task network uses a feature pyramid structure to reuse multi-scale feature maps to improve the pixel semantic recovery accuracy of low-resolution feature maps. Through multi-layer deconvolution operations, the feature vector is mapped to the original image size and the segmentation result is output.

[0081] The task network first upsamples the multi-level features extracted in step 3 and uses the deconvolution kernel to expand the small-sized features. Assuming the input feature size is M1×M1, the convolution kernel size is N×N, the moving step size is s, and the padding size is p, the output feature size o after the deconvolution operation is:

[0082] o=s(M1-1)-2p+N

[0083] After the deconvolution operation, the Relu activation function is connected, and in step 2, the output of the corresponding stage is skipped, that is, the fusion feature matrix and the deconvolution feature matrix of the same dimension size are spliced ​​in the third dimension. Through the skip connection, the bottom-level position information is fused with the deep semantic information. Finally, the feature matrix is ​​gradually expanded through deconvolution kernels of different sizes, and a convolutional neural network with a convolution kernel of 3×3 size is connected to obtain the pixel-level semantic segmentation result map.

[0084] The present embodiment also provides a high-resolution image target recognition and segmentation system, which includes: a first module, used to slice a preset original image into slices of the same size to obtain multiple image slices; perform feature mapping on the multiple image slices through a preset fully connected network to obtain multiple initial vectors; a second module, used to perform feature extraction on the initial vector to obtain a fused feature vector; a third module, used to perform multi-level feature extraction on the fused feature vector to obtain a high-dimensional fused feature vector; and a fourth module, used to upsample the high-dimensional fused feature vector to obtain a semantic segmentation result map.

[0085] This embodiment uses cross-calculation and multi-level use of window self-attention and sliding window self-attention, and uses attention channels to fuse multiple features, so that the network can have good spatial modeling and feature extraction capabilities for high-resolution image target recognition and segmentation tasks, which can significantly improve the accuracy of calculation results.

[0086] This embodiment designs a task network with skip connections. Through the skip connections between the deconvolution layer and the corresponding feature extraction layer, the feature information of the image and the corresponding position information can be fused to provide more accurate information when restoring pixel-level semantic content.

[0087] This embodiment performs two operations, slicing and word embedding layer assignment, on the original high-resolution image. Slicing divides the high-resolution image into small-scale tensors to facilitate network processing and calculation. The word embedding layer adds position association information to each slice, which can increase the network's spatial perception ability and improve network recognition accuracy.

[0088] Based on the modeling of windows and window sliding on images, this embodiment expands the receptive field of the network and enhances the network's feature extraction capability for images based on the multi-directional movement of the window in the image matrix.

[0089] This embodiment uses the attention mechanism to fuse the features extracted by sliding windows in different directions. Through the attention mechanism, the network can learn the correlation information between windows in different directions on the image, which helps to improve the network's recognition accuracy for specific target substructures.

[0090] The mask matrix of this embodiment can separate the associated area and the non-associated area after the window slides by setting the maximum value and the minimum value, so that the features extracted by the network after the window slides are also relevant, avoiding the gradient drift phenomenon during the network training process.

[0091] This embodiment proposes a feature pyramid task network, which consists of a convolutional network layer with input layers ranging from small to large, a maximum pooling layer and a fully connected layer. The extracted high-dimensional features can be restored to a pixel-level semantic segmentation result image of the same shape as the input image.

[0092] Although the present invention has been disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications to the technical solution of the present invention by using the methods and technical contents disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope of the technical solution of the present invention.

Claims

1. A high-resolution image target recognition and segmentation method, characterized in that include: The preset original image is sliced ​​and divided into multiple image slices according to the same size; Multiple image slices are subjected to feature mapping through a preset fully connected network to obtain multiple initial vectors; Perform feature extraction on the initial vector to obtain a fused feature vector; Perform multi-level feature extraction on the fused feature vector to obtain a high-dimensional fused feature vector; The high-dimensional fused feature vector is upsampled to obtain the semantic segmentation result map.

2. The high-resolution image target recognition and segmentation method according to claim 1, characterized in that: The fused feature vector obtained by extracting high-dimensional features from the initial vector includes: The initial vector is normalized and calculated to obtain a normalized vector; a window is established based on the normalized vector; the window is slid on the normalized vector and self-attention is calculated to obtain a sliding window self-attention vector; a mask matrix is ​​calculated on the sliding window self-attention vector to obtain mask vectors in three directions; and attention features are fused on the mask vectors in the three directions to obtain a fused feature vector.

3. The high-resolution image target recognition and segmentation method according to claim 2, characterized in that: The normalized vector includes a mean and a variance.

4. The high-resolution image target recognition and segmentation method according to claim 3, characterized in that: The normalized vector is obtained by the following formula: Among them, μ i is the mean of the i-th initial vector, σ i is the variance of the ith initial vector, W is the number of initial vectors, x i is the i-th initial vector, ε is the noise, and i is the number of the initial vector.

5. The high-resolution image target recognition and segmentation method according to claim 2, characterized in that: The sliding window self-attention vector is obtained by the following formula: Among them, z l+1 is the sliding window self-attention vector, MLP() represents the multi-layer perceptron, WMSA() and SWMSA() represent the window-based self-attention calculation function and the sliding window-based self-attention calculation function respectively, LN() represents the normalization layer calculation function, is the first intermediate vector, z l is the second intermediate vector, is the third intermediate vector, and l is the window number.

6. The high-resolution image target recognition and segmentation method according to claim 2, characterized in that: The three directions are upper left, lower left, and upper right directions for window sliding.

7. The high-resolution image target recognition and segmentation method according to claim 2, characterized in that: The fused feature vector is: Among them, Q is the mask vector in the lower left direction, K is the mask vector in the upper left direction, V is the mask vector in the upper right direction, and d is the scaling factor.

8. A high-resolution image target recognition and segmentation system, characterized in that include: The first module is used to slice and divide the preset original image into multiple image slices of the same size; Multiple image slices are subjected to feature mapping through a preset fully connected network to obtain multiple initial vectors; The second module is used to extract features from the initial vector to obtain a fused feature vector; The third module is used to perform multi-level feature extraction on the fused feature vector to obtain a high-dimensional fused feature vector; The fourth module is used to upsample the high-dimensional fused feature vector to obtain a semantic segmentation result map.

9. The high-resolution image target recognition and segmentation system according to claim 8, characterized in that: The fused feature vector obtained by extracting high-dimensional features from the initial vector includes: The initial vector is normalized and calculated to obtain a normalized vector; a window is established based on the normalized vector; the window is slid on the normalized vector and self-attention is calculated to obtain a sliding window self-attention vector; a mask matrix is ​​calculated on the sliding window self-attention vector to obtain mask vectors in three directions; and attention features are fused on the mask vectors in the three directions to obtain a fused feature vector.

10. The high-resolution image target recognition and segmentation system according to claim 9, characterized in that: The normalized vector includes a mean and a variance.

Images