Image internal occlusion relationship prediction method and system based on edge information guidance

Abstract

The application discloses an image internal occlusion relationship prediction method and system based on edge information guidance. The method comprises two parts of an edge detector and an occlusion relationship predictor, and comprises the following steps: inputting an image through an edge detector composed of an image segmentation module, a plurality of pooling Transformer modules and a deconvolution module to generate an edge detection graph; then in the occlusion relationship predictor, the edge detection graph acts on the image element features generated by the edge detector through an edge information guidance module; then the new image element features are sequentially optimized through a plurality of pooling Transformer modules; and finally, a last deconvolution module adjusts and enlarges the finally obtained image element features in resolution, and outputs an occlusion relationship graph. The application first proposes a new method of guiding feature calculation by the generated edge detection graph and directly generating an occlusion relationship graph. Compared with other existing methods, the application has less calculation amount and better prediction effect.

Description

Technical Field

[0001] This invention belongs to the field of computer vision, and more specifically, relates to a method and system for predicting image internal occlusion relationships based on edge information guidance. Background Technology

[0002] Predicting occlusion relationships within an image using visual algorithms is a crucial technique for visual understanding and analysis. Accurate prediction of occlusion relationships allows for the precise determination of the spatial order of regions within an image using only two-dimensional information, without introducing three-dimensional information. This facilitates the direct extraction of three-dimensional depth information from two-dimensional data, leading to more accurate visual analysis and understanding.

[0003] Currently, the common approach to predicting occlusion relationships within image regions is to use a convolutional neural network to obtain an edge detection map and an occlusion direction map through two separate paths, and then use non-maximum suppression to obtain the final occlusion relationship map. This approach not only involves a cumbersome computational process and inevitably adds the non-maximum suppression step, but also fails to effectively utilize the inherent correlation between the edge detection map and the occlusion relationship map. These problems are detrimental to the effective prediction of the occlusion relationship map. Summary of the Invention

[0004] To address the aforementioned problems and shortcomings of current methods and technologies, this invention proposes for the first time a novel approach and method that uses the generated edge detection map to guide feature computation. This eliminates the need for generating occlusion direction maps and non-maximum suppression processes, enabling the direct generation of occlusion relationship graphs. Compared to other existing methods, this invention boasts a simpler deep neural network structure, requires less computation, and delivers better prediction results.

[0005] The technical solution adopted in this invention is an image internal occlusion relationship prediction method based on edge information, which specifically includes the following steps:

[0006] Step 1: The input image is passed through an edge detector, which consists of an image segmentation module, a pooling Transformer module, and a deconvolution module. The output is an edge detection map.

[0007] Step 2: Input the edge detection map into the occlusion relationship predictor. The edge information guidance module acts on the image meta-features generated by the last pooling Transformer module in the edge detector, so that the edge information can globally guide the generation process of image meta-features and obtain new image meta-features. Then, the new image meta-features are successively passed through several pooling Transformer modules to optimize the image meta-features. Finally, the optimized image meta-features are passed through the deconvolution module to output the occlusion relationship map.

[0008] Preferably, the edge detector includes an image segmentation module, four pooling Transformer modules, and a deconvolution module 1. After the input image is input to the edge detector, it undergoes computation through these modules sequentially. The internal computation and structure of the edge detector specifically include:

[0009] Step 1-1: The input image is processed by an image segmentation module in the edge detector. This module first segments the input image into non-overlapping sub-blocks of size P×P, and stretches the features of each sub-block into an image meta-feature vector. Let the spatial resolution of the input image be H×W, where H and W represent the height and width of the image, respectively. Then, the overall dimension of the image meta-feature obtained after image segmentation and feature stretching is ((H / P)×(W / P))×C1, where C1 is the length of the image meta-feature. Then, the obtained image meta-feature is processed through an MLP (Multilayer Perceptron) layer to adjust the feature dimension, resulting in a new image meta-feature dimension of ((H / P)×(W / P))×C2, where C2 is the length of the image meta-feature after processing by the MLP layer.

[0010] Step 1-2: The image metadata obtained in Step 1-1 is sequentially passed through four cascaded pooling Transformer modules. These four modules have the same structure, consisting of two normalization layers, one feature dimension transformation module 1, one feature dimension transformation module 2, one pooling layer, two summing layers, and one MLP layer. When the image metadata obtained in Step 1-1 is sequentially passed through these four pooling Transformer modules, the dimension of the output image metadata produced by each pooling Transformer module is completely consistent, which is ((H / P)×(W / P))×C3, where C3 is the length of the image metadata output by each module.

[0011] Steps 1-3: The image metadata output from the last pooling Transformer module in Steps 1-2 is fed into the deconvolution module 1. The latter consists of a feature dimension transformation module 1, a 3×3 deconvolution layer, a GELU activation function layer, a BN (batch normalization) layer, and a Softmax function layer connected in series. The feature dimension obtained by the image metadata through the feature dimension transformation module 1 is (H / P)×(W / P)×C4, where C4 is the feature length of the feature map. The output of the deconvolution module 1 is an edge detection map with a spatial size of H×W.

[0012] Preferably, the pooling Transformer module performs the following calculations: First, the input image metadata features are sequentially passed through a normalization layer, feature dimension transformation module 1, a pooling layer, feature dimension transformation module 2, and an addition layer. The normalization layer is implemented through layer normalization calculations. Feature dimension transformation module 1 transforms the image metadata features into a feature map format, adjusting the features from two dimensions (number of metadata × length of metadata) to three dimensions (length of the feature map × width × length of metadata). Then, N×N pooling calculations are performed within the pooling layer, where N is the pooling neighborhood. The pooled feature map is then enlarged to the size before pooling. Next, in the feature dimension conversion module 2, the features are converted into two-dimensional image meta-feature styles. The calculation of the summation layer is achieved by adding the image meta-features. The calculation objects of the summation layer are the input image meta-features and the output image meta-features of the feature dimension conversion module 2. Subsequently, the summed image meta-features are fed into the normalization layer, and then pass through the MLP layer and the summation layer in sequence. The input of the summation layer is the output of the aforementioned summation layer and the output of the MLP layer. Finally, the summed image meta-features are the output of the pooling Transformer module.

[0013] Preferably, the occlusion relationship predictor includes an edge information guidance module, three pooling Transformer modules, and a deconvolution module 2; the internal computation and structure of the occlusion relationship predictor specifically include:

[0014] Step 2-1: Inside the edge information guidance module, the image metadata generated by the last pooling Transformer module of the edge detector is transformed from a two-dimensional feature (number of image metadata × length of image metadata) to a three-dimensional feature (length of feature map × width × length of image metadata) through the feature dimension transformation module 1. Then, the edge detection map is shrunk to the same size as the feature map. Next, the shrunk edge detection map and the feature map are multiplied channel by channel. Finally, the multiplied feature map is transformed into a two-dimensional image metadata feature style through the feature dimension transformation module 2 to obtain the image metadata features guided by edge information.

[0015] Step 2-2: The image meta-features guided by edge information are first passed through three cascaded pooling Transformer modules, and then through a deconvolution module 2, which consists of a feature dimension transformation module 1, two 3×3 deconvolution layers, a GELU activation function layer, a BN (batch normalization) layer, and a Sigmoid function layer, all cascaded together. The feature dimension calculated by the feature dimension transformation module 1 is (H / P)×(W / P)×C5, where C5 is the feature length of the feature map. The output of the deconvolution module 2 is an occlusion relationship map with a spatial size of H×W, which is the final output of this invention.

[0016] Preferably, the present invention trains the entire neural network in an end-to-end manner, and the optimization objective function is: L = L1 + L2, where L1 represents objective function 1, which is applied to the edge detection map, and L1 is a binary cross-entropy function; L2 represents objective function 2, which is applied to the final occlusion relationship map, and it is an orthogonal direction regression loss function.

[0017] Preferably, L1 is calculated as shown in the following formula:

[0018]

[0019] Among them, G - and G + Let a represent the set of non-edge pixels and edge pixels in the manually labeled result, where a = |G + | / (|G - |+|G + |),|G - | and | G + | represent the number of non-edge pixels and edge pixels respectively, 'a' is used to balance the imbalance of sample numbers during edge detection, and 'g'... i This represents the probability value of each pixel belonging to an edge, which is the value of each pixel in the edge detection map;

[0020] The calculation method for L2 is shown in the following formula:

[0021]

[0022] in, This represents the classic smoothing loss. This represents the orientation angle value of the pixel position in the manually labeled result, where x and y are the horizontal and vertical coordinates of the angle of the corresponding pixel position in the generated occlusion relationship map, respectively.

[0023] The present invention also provides an image internal occlusion relationship prediction system based on edge information, comprising the following modules:

[0024] The edge detection map acquisition module is used to pass the input image through the edge detector, which consists of an image segmentation module, a pooling Transformer module, and a deconvolution module, and the output is an edge detection map.

[0025] The occlusion relationship graph acquisition module is used to input the edge detection map into the occlusion relationship predictor. The edge information guidance module acts on the image meta-features generated by the last pooling Transformer module in the edge detector, so that the edge information can globally guide the generation process of image meta-features to obtain new image meta-features. Subsequently, the new image meta-features are then optimized by several pooling Transformer modules. Finally, the optimized image meta-features are passed through the deconvolution module to output the occlusion relationship graph.

[0026] The present invention also provides an electronic device for predicting image internal occlusion relationships based on edge information guidance, comprising:

[0027] One or more processors;

[0028] A storage device is provided for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the image internal occlusion relationship prediction method based on edge information as described above.

[0029] In summary, the technical solutions designed by this invention have the following advantages compared with the prior art:

[0030] This invention is the first to propose a novel approach and method that uses the generated edge detection map to guide feature computation, eliminating the need for generating occlusion direction maps and non-maximum suppression processes, and directly generating occlusion relationship graphs. Compared to existing methods, this invention boasts a simpler deep neural network structure, requires less computation, and achieves better prediction results. Attached Figure Description

[0031] Figure 1 This invention provides an overall flowchart of an image internal occlusion relationship prediction method based on edge information.

[0032] Figure 2 This is a structural diagram of the pooled Transformer module;

[0033] Figure 3 This is a schematic diagram of the edge information guidance module. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0035] This invention provides a method for predicting image internal occlusion relationships based on edge information, such as... Figure 1 As shown, it includes the following steps:

[0036] Step 1: The input image is passed through an edge detector, which consists of an image segmentation module, a pooling Transformer module, and a deconvolution module; the output of the edge detector is an edge detection map.

[0037] Step 2: In the occlusion relationship predictor, the edge detection map is applied to the image meta-features generated by the last pooling Transformer module in the edge detector through the edge information guidance module, so that the edge information can globally guide the generation process of the image meta-features; then the new image meta-features are optimized by several pooling Transformer modules in sequence; finally, the generated image meta-features are passed through the deconvolution module to output the occlusion relationship map.

[0038] The invention mainly consists of two components: (1) an edge detector and (2) an occlusion relationship predictor. The steps in this invention will be described in detail below.

[0039] 1. Implementation of edge detectors

[0040] The edge detector in this invention includes an image segmentation module, four pooling Transformer modules, and a deconvolution module 1, as shown below. Figure 1 As shown. After the input image is input to the edge detector, it undergoes calculations sequentially through these modules; the internal calculations and structure of the edge detector specifically include:

[0041] Step 1-1: The input image is processed by an image segmentation module in the edge detector. This module first segments the input image into non-overlapping sub-blocks of size P×P (in this example, P is 4), and stretches the features of each sub-block into an image meta-feature vector. Let the spatial resolution of the input image be H×W, where H and W represent the height and width of the image, respectively. Then, the overall dimension of the image meta-feature obtained after image segmentation and feature stretching is ((H / P)×(W / P))×C1, where C1 is the length of the image meta-feature (in this example, C1 is 32). Then, the obtained image meta-feature is processed through an MLP (Multilayer Perceptron) layer to adjust the feature dimension, resulting in a new image meta-feature dimension of ((H / P)×(W / P))×C2, where C2 is the length of the image meta-feature after processing by the MLP layer (in this example, C2 is 64).

[0042] Step 1-2: The image metadata obtained in Step 1-1 is sequentially passed through four cascaded pooling Transformer modules. These four modules have the same structure, consisting of two normalization layers, a feature dimension transformation module 1, a feature dimension transformation module 2, a pooling layer, two summing layers, and an MLP layer. When the image metadata obtained in Step 1-1 is sequentially passed through these four pooling Transformer modules, the dimension of the output image metadata generated by each pooling Transformer module is completely consistent, which is ((H / P)×(W / P))×C3, where C3 is the length of the image metadata output by each module. In this embodiment, the value of C3 is 64.

[0043] Steps 1-3: The image metadata output from the last pooling Transformer module in Steps 1-2 is fed into the deconvolution module 1. The latter consists of a feature dimension conversion module 1, a 3×3 deconvolution layer, a GELU activation function layer, a BN (batch normalization) layer, and a Softmax function layer connected in series. The feature dimension obtained by the image metadata through the feature dimension conversion module 1 is (H / P)×(W / P)×C4, where C4 is the feature length of the feature map. In this embodiment, C4 is 128. The output of the deconvolution module 1 is an edge detection map with a spatial size of H×W.

[0044] In steps 1-2, the pooling Transformer module is used, such as... Figure 2 As shown, the specific calculation is as follows: First, the input image metadata features are sequentially passed through a normalization layer, feature dimension conversion module 1, a pooling layer, feature dimension conversion module 2, and an addition layer; wherein, the normalization layer is implemented through layer normalization calculation; feature dimension conversion module 1 transforms the image metadata features into the style of a feature map, that is, it adjusts the features with two dimensions (number of image metadata × length of image metadata) to features with three dimensions (length of feature map × width × length of image metadata); then, N×N pooling calculation is performed in the pooling layer, where N is the neighborhood range of pooling. In this embodiment, the value of N is 9, and the pooling... The feature map after pooling is enlarged to the size before pooling; then, in the feature dimension conversion module 2, the features are converted into two-dimensional image meta-feature styles; the calculation of the summation layer is achieved by adding the image meta-features, and the calculation objects of the summation layer are the input image meta-features and the output image meta-features of the feature dimension conversion module 2; subsequently, the summed image meta-features are fed into the normalization layer, and then pass through the MLP layer and the summation layer in sequence, with the input of the summation layer being the output of the aforementioned summation layer and the output of the MLP layer; finally, the summed output image meta-features are the output result of the pooling Transformer module.

[0045] 2. Implementation of the Occlusion Relationship Predictor

[0046] The occlusion relationship predictor in this invention includes an edge information guidance module, three pooling Transformer modules, and a deconvolution module 2; the internal computation and structure of the occlusion relationship predictor specifically include:

[0047] Step 2-1, within the edge information guidance module, such as Figure 3 As shown, the image metadata generated by the last pooling Transformer module of the edge detector is transformed from a two-dimensional feature (number of pixels × length of pixels) to a three-dimensional feature (length of feature map × width × length of pixels) through the feature dimension transformation module 1. Then, the edge detection map is shrunk to the same size as the feature map. Next, the shrunk edge detection map and the feature map are multiplied channel by channel. Finally, the multiplied feature map is transformed into a two-dimensional image metadata feature style through the feature dimension transformation module 2 to obtain the image metadata features guided by edge information.

[0048] Step 2-2: The image meta-features guided by edge information are first passed through three cascaded pooling Transformer modules, and then through a deconvolution module 2, which consists of a feature dimension transformation module 1, two 3×3 deconvolution layers, a GELU activation function layer, a BN (batch normalization) layer, and a Sigmoid function layer, all cascaded together. The feature dimension calculated by the feature dimension transformation module 1 is (H / P)×(W / P)×C5, where C5 is the feature length of the feature map. In this embodiment, C5 is 128. The output of the deconvolution module 2 is an occlusion relationship map with a spatial size of H×W, which is the final output of this invention.

[0049] This invention trains the entire neural network in an end-to-end manner. The objective function for optimization is: L = L1 + L2, where L1 represents objective function 1, which is applied to the edge detection map and is a binary cross-entropy function; L2 represents objective function 2, which is applied to the final occlusion relationship map and is an orthogonal regression loss function.

[0050] The calculation method for L1 is shown in the following formula:

[0051]

[0052] Among them, G - and G + Let a represent the set of non-edge pixels and edge pixels in the manually labeled result, where a = |G + | / (|G - |+|G + |),|G - | and | G +| represent the number of non-edge pixels and edge pixels respectively, 'a' is used to balance the imbalance of sample numbers during edge detection, and 'g'... i This represents the probability value of each pixel belonging to an edge, which is the value of each pixel in the edge detection map.

[0053] The calculation method for L2 is shown in the following formula:

[0054]

[0055] in, This represents the classic smoothing loss. This represents the orientation angle value of the pixel position in the manually labeled result, where x and y are the horizontal and vertical coordinates of the angle of the corresponding pixel position in the generated occlusion relationship map, respectively.

[0056] The table below shows the results of comparing the present invention with two mainstream methods, DOOBNet and MTORL, on the mainstream PIOD database using four common metrics: B-ODS, B-OIS, O-ODS, and O-OIS. As can be seen from the table, the present invention outperforms the two comparison methods in all four metrics, demonstrating the effectiveness and advantages of the present invention.

[0057]

[0058]

[0059] The present invention also provides an image internal occlusion relationship prediction system based on edge information, comprising the following modules:

[0060] The edge detection map acquisition module is used to pass the input image through the edge detector, which consists of an image segmentation module, a pooling Transformer module, and a deconvolution module, and the output is an edge detection map.

[0061] The occlusion relationship graph acquisition module is used to input the edge detection map into the occlusion relationship predictor. The edge information guidance module acts on the image meta-features generated by the last pooling Transformer module in the edge detector, so that the edge information can globally guide the generation process of image meta-features to obtain new image meta-features. Subsequently, the new image meta-features are then optimized by several pooling Transformer modules. Finally, the optimized image meta-features are passed through the deconvolution module to output the occlusion relationship graph.

[0062] The specific implementation methods of each module are the same as those of each step, and will not be described in this invention.

[0063] The present invention also provides an electronic device for predicting image internal occlusion relationships based on edge information guidance, comprising:

[0064] One or more processors;

[0065] A storage device is provided for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the image internal occlusion relationship prediction method based on edge information as described above.

[0066] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting image occlusion relationships based on edge information, characterized in that, Includes the following steps: Step 1: The input image is passed through an edge detector, which consists of an image segmentation module, a pooling Transformer module, and a deconvolution module. The output is an edge detection map. The edge detector includes an image segmentation module, four pooling Transformer modules, and a deconvolution module. After the input image is fed into the edge detector, it undergoes sequential computation through these modules. The internal computation and structure of the edge detector specifically include: Step 1-1: The input image is processed by an image segmentation module in the edge detector. The image segmentation module first segments the input image into non-overlapping sub-blocks of size P×P, and stretches the features of each sub-block into an image meta-feature vector. Let the spatial resolution of the input image be H×W, where H and W represent the height and width of the image, respectively. Then, the overall dimension of the image meta-feature obtained after image segmentation and feature stretching is ((H / P)×(W / P))×C1, where C1 is the length of the image meta-feature. Then, the obtained image meta-feature is processed by a multilayer perceptron (MLP) layer to adjust the feature dimension, resulting in a new image meta-feature dimension of ((H / P)×(W / P))×C2, where C2 is the length of the image meta-feature after MLP processing. Step 1-2: The image metadata obtained in Step 1-1 is sequentially passed through four cascaded pooling Transformer modules. These four modules have the same structure, consisting of two normalization layers, one feature dimension transformation module 1, one feature dimension transformation module 2, one pooling layer, two summing layers, and one MLP layer. When the image metadata obtained in Step 1-1 is sequentially passed through these four pooling Transformer modules, the dimension of the output image metadata produced by each pooling Transformer module is completely consistent, which is ((H / P)×(W / P))×C3, where C3 is the length of the image metadata output by each module. Steps 1-3: The image metadata output from the last pooling Transformer module in Steps 1-2 is fed into the deconvolution module 1, and the output is an edge detection map with spatial dimensions of H×W. Step 2: The edge detection map is input into the occlusion relationship predictor. The edge information guidance module is applied to the image meta-features generated by the last pooling Transformer module in the edge detector, so that the edge information can globally guide the generation process of image meta-features to obtain new image meta-features. Then, the new image meta-features are sequentially passed through several pooling Transformer modules to optimize the image meta-features. Finally, the optimized image meta-features are passed through the deconvolution module to output the occlusion relationship map. The occlusion relationship predictor includes an edge information guidance module, three pooling Transformer modules, and one deconvolution module. The internal computation and structure of the occlusion relationship predictor specifically include: Step 2-1: Inside the edge information guidance module, the image meta-features generated by the last pooling Transformer module of the edge detector are adjusted from two-dimensional features to three-dimensional features through the feature dimension conversion module 1. Then, the edge detection map is shrunk to the same size as the feature map. Next, the shrunk edge detection map and the feature map are multiplied channel by channel. Finally, the multiplied feature map is converted into a two-dimensional image meta-feature style through the feature dimension conversion module 2 to obtain the image meta-features guided by edge information. Step 2-2: The image meta-features guided by edge information are first passed through three cascaded pooling Transformer modules, and then through deconvolution module 2. The output is an occlusion relationship map with a spatial size of H×W, which is the final output result.

2. The image occlusion relationship prediction method based on edge information as described in claim 1, characterized in that: The deconvolution module 1 consists of a feature dimension transformation module 1, a 3×3 deconvolution layer, a GELU activation function layer, a batch normalization layer, and a Softmax function layer connected in series. The feature dimension of the image meta-features calculated by the feature dimension transformation module 1 is (H / P)×(W / P)×C4, where C4 is the feature length of the feature map.

3. The image internal occlusion relationship prediction method based on edge information as described in claim 1, characterized in that: The specific calculation of the pooling Transformer module is as follows: First, the input image metadata features are sequentially passed through a normalization layer, feature dimension conversion module 1, a pooling layer, feature dimension conversion module 2, and an addition layer. The normalization layer is implemented through layer normalization calculation. Feature dimension conversion module 1 transforms the image metadata features into a feature map style, adjusting the dimensions from two dimensions to three dimensions. Then, N×N pooling is performed in the pooling layer, where N is the pooling neighborhood, and the pooled feature map is enlarged to the size before pooling. Next, in feature dimension conversion module 2, the features are converted back to a two-dimensional image metadata feature style. The addition layer is implemented by adding the image metadata features, with the input image metadata features and the output image metadata features from feature dimension conversion module 2 as the calculation objects. Subsequently, the added image metadata features are fed into the normalization layer, and then sequentially passed through an MLP layer and the addition layer. The input to the addition layer is the output of the aforementioned addition layer and the output of the MLP layer. Finally, the added image metadata features are the output result of the pooling Transformer module.

4. The image occlusion relationship prediction method based on edge information as described in claim 1, characterized in that: The deconvolution module 2 consists of a feature dimension transformation module 1, two 3×3 deconvolution layers, a GELU activation function layer, a batch normalization layer, and a Sigmoid function layer connected in series. The feature dimension of the image meta-features calculated by the feature dimension transformation module 1 is (H / P)×(W / P)×C5, where C5 is the feature length of the feature map.

5. The image occlusion relationship prediction method based on edge information as described in claim 1, characterized in that: The entire neural network consisting of steps 1 and 2 is trained in an end-to-end manner. The objective function for optimization is: L = L1 + L2, where L1 represents objective function 1, which is applied to the edge detection map and is a binary cross-entropy function; L2 represents objective function 2, which is applied to the final occlusion relationship map and is an orthogonal regression loss function.

6. The image internal occlusion relationship prediction method based on edge information as described in claim 5, characterized in that: The calculation method for L1 is shown in the following formula: ； in, and This represents the set of non-edge pixels and edge pixels in the manually labeled result. , and These represent the number of non-edge pixels and the number of edge pixels, respectively. Used to balance the imbalance of sample numbers in edge detection. This represents the probability value of each pixel belonging to an edge, which is the value of each pixel in the edge detection map; The calculation method for L2 is shown in the following formula: ； in, This represents the classic smoothing loss. This represents the orientation angle value of the pixel position in the manually labeled result. and These are the horizontal and vertical coordinates of the angle of the corresponding pixel position in the generated occlusion relationship map.

7. A system for predicting image internal occlusion relationships based on edge information, characterized in that, Includes the following modules: The edge detection map acquisition module is used to pass the input image through the edge detector, which consists of an image segmentation module, a pooling Transformer module, and a deconvolution module, and the output is an edge detection map. The edge detector includes an image segmentation module, four pooling Transformer modules, and a deconvolution module. After the input image is fed into the edge detector, it undergoes sequential computation through these modules. The internal computation and structure of the edge detector specifically include: Step 1-1: The input image is processed by an image segmentation module in the edge detector. The image segmentation module first segments the input image into non-overlapping sub-blocks of size P×P, and stretches the features of each sub-block into an image meta-feature vector. Let the spatial resolution of the input image be H×W, where H and W represent the height and width of the image, respectively. Then, the overall dimension of the image meta-feature obtained after image segmentation and feature stretching is ((H / P)×(W / P))×C1, where C1 is the length of the image meta-feature. Then, the obtained image meta-feature is processed by a multilayer perceptron (MLP) layer to adjust the feature dimension, resulting in a new image meta-feature dimension of ((H / P)×(W / P))×C2, where C2 is the length of the image meta-feature after MLP processing. Step 1-2: The image metadata obtained in Step 1-1 is sequentially passed through four cascaded pooling Transformer modules. These four modules have the same structure, consisting of two normalization layers, one feature dimension transformation module 1, one feature dimension transformation module 2, one pooling layer, two summing layers, and one MLP layer. When the image metadata obtained in Step 1-1 is sequentially passed through these four pooling Transformer modules, the dimension of the output image metadata produced by each pooling Transformer module is completely consistent, which is ((H / P)×(W / P))×C3, where C3 is the length of the image metadata output by each module. Steps 1-3: The image metadata output from the last pooling Transformer module in Steps 1-2 is fed into the deconvolution module 1, and the output is an edge detection map with spatial dimensions of H×W. The occlusion relationship graph acquisition module is used to input the edge detection map into the occlusion relationship predictor. The edge information guidance module acts on the image meta-features generated by the last pooling Transformer module in the edge detector, so that the edge information can globally guide the generation process of image meta-features to obtain new image meta-features. Subsequently, the new image meta-features are then optimized by several pooling Transformer modules. Finally, the optimized image meta-features are passed through the deconvolution module to output the occlusion relationship graph. The occlusion relationship predictor includes an edge information guidance module, three pooling Transformer modules, and one deconvolution module. The internal computation and structure of the occlusion relationship predictor specifically include: Step 2-1: Inside the edge information guidance module, the image meta-features generated by the last pooling Transformer module of the edge detector are adjusted from two-dimensional features to three-dimensional features through the feature dimension conversion module 1. Then, the edge detection map is shrunk to the same size as the feature map. Next, the shrunk edge detection map and the feature map are multiplied channel by channel. Finally, the multiplied feature map is converted into a two-dimensional image meta-feature style through the feature dimension conversion module 2 to obtain the image meta-features guided by edge information. Step 2-2: The image meta-features guided by edge information are first passed through three cascaded pooling Transformer modules, and then through deconvolution module 2. The output is an occlusion relationship map with a spatial size of H×W, which is the final output result.

8. An electronic device for predicting image internal occlusion relationships based on edge information guidance, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement an image internal occlusion relationship prediction method based on edge information as described in any one of claims 1 to 6.

Images