Unsupervised multi-view stereo reconstruction method for occlusion-resistant regions

Abstract

The present invention discloses an unsupervised multi-view stereo reconstruction method for anti-occlusion areas, which can effectively solve the three-dimensional reconstruction problem of occluded areas by processing images from multiple perspectives without relying on real depth information, and calculate a point cloud model with high precision. First, multi-scale features are extracted; then, a deformable large-core attention network is used to aggregate global information, and the depth map and corresponding confidence map of the scene are calculated; secondly, the contrast consistency loss and the perceptual consistency loss are calculated to provide a pseudo-supervision signal for the depth estimation process and optimize the depth map; finally, the depth map is fused to obtain a point cloud model. The present invention makes full use of the unsupervised multi-view stereo calculation method based on multiple supervision items, effectively mines the features of the input image itself, estimates a high-quality depth map, and then calculates an accurate point cloud model.

Description

Technical Field

[0001] The present invention relates to multi-view Figure 3 The present invention relates to the field of three-dimensional reconstruction, computer graphics and computer vision technology, and specifically to an unsupervised multi-view stereo reconstruction method for anti-occlusion areas. Background Art

[0002] Multi-view Stereo (MVS) is a technique for reconstructing a point cloud model of a scene from multi-view images. Based on the data format of the output reconstruction results, 3D reconstruction methods can generally be divided into the following categories: (1) direct point cloud reconstruction; (2) voxel-based methods; (3) grid-based methods; and (4) methods based on depth map fusion.

[0003] The MVS method based on supervised learning requires ground truth as training data. However, the process of obtaining large-scale ground truth data is very complicated, which limits the generalization ability of such methods in unknown environments. The emergence of unsupervised methods solves this problem. Such methods do not require ground truth as supervision information and can reconstruct three-dimensional point cloud models based on multi-view images, which significantly improves their adaptability in different scenes. Existing unsupervised methods are mainly based on the photometric consistency assumption, that is, when observing the same scene point in different perspectives or adjacent frames, it is assumed that its photometric value (such as grayscale value or color) remains consistent when changing. However, this assumption often fails in low-texture areas or occluded areas. Low-texture areas lack sufficient photometric information for matching, while in occluded areas, some parts of the scene may be occluded by other objects in some or all perspectives, resulting in matching errors. Occlusion is particularly common in complex scenes, such as urban street scenes or indoor environments.

[0004] The above problems seriously hinder the development and application of unsupervised multi-view stereo reconstruction methods, and expose the shortcomings of existing unsupervised multi-view stereo reconstruction methods when processing image data in occluded areas. Therefore, people are in urgent need of a robust and high-precision stereo reconstruction method to effectively restore the depth value of images in occluded areas.

[0005] In recent years, there have been relevant research papers to further improve the accuracy of unsupervised multi-view stereo reconstruction. However, these existing unsupervised MVS methods usually rely on the photometric consistency assumption to estimate the depth of the three-dimensional scene. However, when there is occlusion in the scene, this assumption does not hold, resulting in the unreliability of the self-supervisory signal. This unreliability will further damage the final reconstruction performance.

[0006] That is to say, existing unsupervised multi-view stereo reconstruction methods still face the following challenges when dealing with depth estimation in occluded areas: (1) Existing unsupervised multi-view stereo reconstruction methods are based on the photometric consistency assumption. However, in occluded areas, the photometric information of the same scene point under different perspectives cannot be kept consistent, resulting in low reconstruction accuracy and unable to meet application requirements in complex scenes; (2) Existing unsupervised multi-view stereo reconstruction methods lack effective occlusion detection and processing mechanisms when dealing with occluded areas, resulting in incorrect depth estimation, especially when there are large-area occluders or dynamic objects in the scene, the depth estimation results are often unstable; (3) Existing unsupervised multi-view stereo reconstruction methods rely on a single supervisory signal when dealing with occluded areas, and it is difficult to provide sufficient optimization information without real depth data, which affects the reconstruction efficiency and the accuracy of the final point cloud model. Summary of the invention

[0007] Purpose of the invention: The purpose of the present invention is to solve the deficiencies in the prior art and to provide an unsupervised multi-view stereo reconstruction method for anti-occlusion areas.

[0008] Technical solution: An unsupervised multi-view stereoscopic reconstruction method for an anti-occlusion region of the present invention comprises the following steps:

[0009] Step 1: Input multi-view image data and camera parameters ; In A multi-view image includes a reference view and Source View ;

[0010] in, express The serial number of the multi-view image in, ; express The number of multi-view images in ; represents the internal parameter matrix of the kth camera, ; and Respectively represent the first camera and the The rotation matrix and translation matrix between the cameras;

[0011] Step 2: For In Image , a feature pyramid network FPN consisting of three stages is used to calculate the multi-scale features of multi-view images; in order to enhance FPN's attention to important features, the channel attention (CA) and spatial attention (SA) modules are introduced to improve its ability to identify key information;

[0012] At each stage , both extracted using feature pyramid network FPN After completing the three stages, we get the multi-view image feature map. ;in, ;

[0013] Step 3: Based on the input multi-view image A series of depth hypothesis planes are generated from the known depth range, and the depth hypothesis planes are evenly spaced in the interval between the minimum depth value and the maximum depth value, and all depth hypothesis planes are extended to match the input multi-view image ( ) of the same size;

[0014] Step 4: Use differentiable homography to transform the multi-view image features obtained in step 2 Projected onto the depth hypothesis plane based on the expansion in step 3, a set of reprojected features is obtained ; Then reproject the features With reference view Corresponding features Compare to construct the cost body;

[0015] Step 5: Based on the cost body obtained in step 4, the deformable large core attention network DLA-Net with global information normalization module is used to aggregate the cost body to obtain a probability body, and the probability body is normalized to obtain an initial depth map;

[0016] Repeat steps 2 to 5 to obtain the final depth map;

[0017] Step 6: During the unsupervised learning training process, photometric consistency loss, contrast consistency loss, and perceptual consistency loss are used to optimize the final depth map obtained in step 5.

[0018] Step 7: Repeat steps 2 to 6 to calculate the depth maps of all perspectives, fuse the depth maps of all perspectives, and output a globally consistent 3D point cloud reconstruction result.

[0019] Furthermore, in step 2, Feature maps of multi-view images The resolution is ;in, Indicates the width of the multi-view image; Indicates the height of the multi-view image; Indicates the number of stages.

[0020] Furthermore, in the step 5, the deformable large core attention module first extracts preliminary features through 3D convolution, and passes through an activation function to enhance nonlinear characteristics. Then, two layers of deformable deep convolution are used to capture spatial offset features. Subsequently, a layer of 3D convolution is used to fuse the features of multiple branches, and residual connections are used to retain input information while enhancing the expression ability of the model. Finally, another layer of 3D convolution is used to further fuse features. The specific contents are as follows:

[0021] Step 5.1: First, use 3D convolution to extract features from the cost volume obtained in step 4. Then, use 3D deconvolution to upsample the feature map. Then, combine the upsampling with the deformable large-core attention module, and use deconvolution again to gradually restore the spatial resolution of the feature map. After the above steps, the regularized cost volume is obtained. ;

[0022] Finally, after a layer The convolution operation of the deformable large core attention network outputs the normalized probability volume , the probability volume represents the matching degree of pixels under different depth assumptions; the normalization process can be expressed as: ;

[0023] in, Depth Hypothesis Plane The probability body under Represents the number of depth hypothesis planes at the coarsest stage , in the middle stage , in the final refinement stage ; The function is used to convert the cost body into a probability body;

[0024] Step 5.2: Probability body Perform depth regression to obtain a depth map :

[0025] ;

[0026] in, Represents a deep hypothetical planar graph; Represents the number of deep hypothesis graphs at the coarsest stage , in the middle stage , in the final refinement stage ; Depth Hypothesis Plane The probability body below.

[0027] Furthermore, the photometric consistency loss function expression in step 6 is as follows:

[0028] ;

[0029] in, represents the reference view based on the input in step S1, Indicates An unwarped image, Indicates the sequence number of the dewarped image, ; Indicates the number of multi-view images; Represents a binary significance mask.

[0030] Furthermore, the step 6 constructs an unsupervised contrast consistency loss strategy, and the detailed steps of providing a pseudo-supervisory signal are as follows:

[0031] First, a center point is randomly selected in the image and the size is set. A connected region is generated based on the connected component analysis algorithm, while ensuring that the size is within the preset range and the generated region does not exceed the image boundary. Subsequently, a dilation operation is applied to ensure the continuity of the connected region. The generated connected region is applied to the initial occlusion mask, and the pixels in the region are marked as 0, indicating that they are occluded, and the number of occluded pixels is dynamically updated. Finally, the connected region is marked as the occluded region, and the rest is marked as the unoccluded region, thereby generating a block-structured occlusion mask.

[0032] Multiply the multi-view images by the corresponding structured occlusion masks element by element to ensure that the unoccluded parts are fully preserved, thereby generating a set of enhanced samples ,in, express The serial number of the multi-view image in, ; Indicates the number of views.

[0033] Input the multi-view image data into step 1 As a regular sample, combined with enhanced samples , both of them can obtain the depth map of the conventional sample through the processing flow of steps S2 to S5 And the depth map of the enhanced sample , calculate the contrast consistency loss between the two , providing a pseudo-supervisory signal for the depth estimation process:

[0034] ;

[0035] in, Indicates the sequence number of the multi-view image, ; represents a binary validity mask; Indicates the number of views.

[0036] Furthermore, the step 6 constructs a perceptual consistency loss strategy, and the detailed steps of providing a pseudo-supervision signal are as follows:

[0037] Using the pre-trained feature extraction network VGG19, extract the multi-view image data based on the input in step S1 in the VGG19 specific layers ('relu_1_1', 'relu_2_1', 'relu_3_1', 'relu_4_1') Medium Reference View High-level semantic features ;

[0038] Fine-tune the feature extraction network VGG19, add a dropout layer at the end of the network, and extract the undistorted image in a specific layer High-level semantic features ,in, Indicates The high-level semantic features of the unwarped image. ; Indicates the number of views; the inverse warped image refers to the intermediate image generated when the final depth map is optimized using the photometric consistency loss function;

[0039] Using perceptual consistency loss to enforce feature semantic similarity between the reference image and the inverse warped image provides an effective pseudo-supervision signal for model learning:

[0040] ;

[0041] in, represents the weight of the perceptual consistency loss. According to experience, ; Indicates the sequence number of the multi-view image, ; Indicates the number of views; Indicates High-level semantic features of the unwarped image.

[0042] Beneficial effects: Compared with the prior art, the advantages of the present invention are as follows:

[0043] (1) The present invention provides an unsupervised multi-view stereo (MVS) method that can calculate a high-quality point cloud model of an occluded scene from multi-view images without the need for true depth.

[0044] (2) The present invention adopts a deformable large-core attention network to overcome the limitations of convolutional neural networks in global information integration and significantly improves the network's ability to handle scene occlusion.

[0045] (3) The present invention adopts a structured occlusion module to construct enhanced samples (positive samples), which are used for contrastive learning. This method can effectively help the model obtain more complete depth estimation in indistinguishable areas, especially when dealing with occlusion problems, thereby improving the accuracy of the overall prediction.

[0046] (4) The present invention adopts a perceptual consistency module, which can still provide an effective pseudo-supervision signal when the photometric consistency assumption fails, thereby effectively guiding the depth recovery task and improving the accuracy of depth estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0048] Figure 2 This is the overall network structure diagram of the present invention.

[0049] Figure 3 is a sample of multi-view image data in an embodiment.

[0050] Figure 4 is a sample of the depth map in the embodiment.

[0051] Figure 5 It is the point cloud model finally output by the embodiment.

[0052] Figure 6 Point cloud model output by existing methods.

[0053] Figure 7 This is the point cloud model output by the existing supervised method CasMVSNet.

[0054] Figure 8 It is the point cloud model output by the existing unsupervised method RC-MVSNet.

[0055] Fig. 9 This is the point cloud model output by the technical solution of the present invention. DETAILED DESCRIPTION

[0056] The technical solution of the present invention is described in detail below, but the protection scope of the present invention is not limited to the embodiments.

[0057] like Figure 1 As shown, an unsupervised multi-view stereoscopic reconstruction method for an anti-occlusion region of the present invention comprises the following steps:

[0058] Step S1: Input multi-view image data and camera parameters ; In A multi-view image includes a reference view and Source View ;

[0059] in, express The serial number of the multi-view image in, ; express The number of multi-view images in ; represents the internal parameter matrix of the kth camera, ; and Respectively represent the first camera and the The rotation matrix and translation matrix between the cameras;

[0060] Step S2: In Image , a feature pyramid network FPN consisting of three stages is used to calculate the multi-scale features of multi-view images; the feature pyramid network FPN introduces a channel attention module CA and a spatial attention module SA;

[0061] Each stage All use the above feature pyramid network FPN extraction After completing the three stages, we get the multi-view image feature map. ; ;

[0062] Step S3: Based on the input multi-view image A series of depth hypothesis planes are generated from the known depth range, and the depth hypothesis planes are evenly spaced in the interval between the minimum depth value and the maximum depth value, and all depth hypothesis planes are extended to match the input multi-view image ( ) of the same size;

[0063] Step S4: Use differentiable homography to transform the multi-view image features obtained in step S2 Projected onto the depth hypothesis plane based on the expansion in step S3, a set of reprojected features is obtained ; Then reproject the features With reference view Corresponding features Compare to construct the cost body;

[0064] Step S5: Based on the cost body obtained in step S4, the cost body is aggregated using the deformable large core attention network DLA-Net of the global information normalization module to obtain a probability body, and the probability body is normalized to obtain an initial depth map;

[0065] Repeat steps S2 to S5 to obtain a final depth map;

[0066] Step S6, in the unsupervised learning training process, using photometric consistency loss, contrast consistency loss and perceptual consistency loss to optimize the final depth map obtained in step S5;

[0067] Step S7, repeating steps S2 to S6, calculating the depth maps of all viewing angles, fusing the depth maps of all viewing angles, and outputting a globally consistent 3D point cloud reconstruction result.

[0068] In step S2 of this embodiment, Feature maps of multi-view images The resolution is ;in, Indicates the width of the multi-view image; Indicates the height of the multi-view image; Indicates the number of stages.

[0069] The present invention builds a model based on a deformable large-core attention network. The deformable large-core attention network enhances the ability to aggregate global information through an adaptive attention mechanism. Especially in the presence of complex occlusions in the scene, it can more effectively capture remote dependency information and alleviate the problem of missing local information. Secondly, a structured occlusion module based on connected component analysis simulates common occlusion scenes in the real world and generates reliable positive samples by accurately modeling the occluded areas; these positive samples are used for contrastive learning to improve the robustness of the model when dealing with occlusion. Finally, perceptual consistency loss provides a new pseudo-supervision signal for the prediction of depth maps. It uses high-level semantic features to enable depth prediction to get rid of the dependence on precise photometric matching. Through this high-level perceptual supervision, the model can more accurately predict depth information when dealing with complex scenes.

[0070] The detailed steps of the above step S5 are:

[0071] Step S5.1, the deformable large core attention module first extracts preliminary features through 3D convolution and passes through an activation function to enhance nonlinear characteristics. Then, two layers of deformable deep convolution are used to capture spatial offset features. Subsequently, a layer of 3D convolution is used to fuse the features of multiple branches and residual connections are used. Finally, another layer of 3D convolution is used to further fuse the features.

[0072] Step S5.2: First, use 3D convolution to extract features from the cost volume obtained in step S4. Then, use 3D deconvolution to upsample the feature map. Then, combine the upsampling with the deformable large-core attention module in step S5.1, and use deconvolution again to gradually restore the spatial resolution of the feature map. After the above steps, the regularized cost volume is obtained. ;

[0073] Finally, after a layer The convolution operation of the deformable large core attention network outputs the normalized probability volume , the probability volume represents the matching degree of pixels under different depth assumptions; the normalization formula is: ;

[0074] in, Depth Hypothesis Plane The probability body under represents the number of depth hypothesis planes;

[0075] Step S5.3: Probability body Perform depth regression to obtain a depth map :

[0076] .

[0077] The above step S6 constructs an unsupervised contrast consistency loss function and provides a pseudo-supervisory signal in detail as follows:

[0078] First, a center point is randomly selected in the image and its size is set. A connected region is generated based on the connected component analysis algorithm, and the connected region does not exceed the image boundary. Then, the connected region generated by the dilation operation is applied to the initial occlusion mask, and the pixels in the region are marked as 0, indicating that they are occluded, and the number of occluded pixels is dynamically updated. Finally, the connected region is marked as the occluded region, and the rest is marked as the unoccluded region, thereby generating a block-structured occlusion mask.

[0079] Multiply the multi-view images by the corresponding structured occlusion masks element-wise to generate a set of enhanced samples ;

[0080] The multi-view image data of step S1 As a regular sample, combined with enhanced samples , both are processed by steps S2 to S5 to obtain the depth map of the conventional sample And the depth map of the enhanced sample ;

[0081] Finally, the contrast consistency loss between the two types of depth maps is calculated , providing a pseudo-supervisory signal for the depth estimation process:

[0082] ;

[0083] Represents a binary significance mask.

[0084] The detailed steps of constructing the perceptual consistency loss function in step S6 above are as follows:

[0085] First, use the pre-trained feature extraction network VGG19 to extract the reference view High-level semantic features ;

[0086] Fine-tune the feature extraction network VGG19, that is, add a discard layer at the end of the feature extraction network VGG19 to extract the inverse distorted image obtained in step S6 High-level semantic features ;

[0087] in, Indicates High-level semantic features of the unwarped image;

[0088] Use perceptual consistency loss to enforce feature semantic similarity between the reference image and the unwarped image , providing effective pseudo-supervision signals for model learning:

[0089] ;

[0090] in, The weight representing the perceptual consistency loss.

[0091] It can be seen from the above embodiments that the overall network structure diagram of the present invention is as follows: Figure 2 As shown, the structured occlusion module, pixel-level perceptual consistency and deformable large-core attention network are fully utilized to effectively solve the limitations of existing unsupervised multi-view stereo reconstruction methods in dealing with occlusion and texture missing areas, and then calculate an accurate point cloud model. The present invention first extracts multi-scale features of multi-view images; then, a deformable large-core attention network is used to aggregate global information and calculate the depth map and confidence map of the reference view; secondly, the contrast consistency loss and perceptual consistency loss are calculated to provide pseudo-supervisory signals for the depth estimation process and optimize the depth map; finally, the depth map is fused to obtain the point cloud model.

[0092] In this embodiment, the input image data sample is as follows: Figure 3 As shown, Figure 3 There are four images from different perspectives. The point cloud model output by this embodiment is as follows: Figure 5As shown, it can be seen that the cloud model reconstructed from the image data has a high geometric consistency with the real scene.

[0093] according to Fig. 9 The final experimental results shown; for Figure 3 The same sample data in Figure 7 and Figure 8 The existing technology is processed in the figure. From the comparison results, it can be seen that the present invention can better restore the detailed structure of the building, especially in the texture details and shadow processing of the roof. Compared with other methods, the present invention reduces structural blur and detail loss, ensuring higher reconstruction accuracy.

Claims

1. An unsupervised multi-view stereo reconstruction method for anti-occlusion regions, characterized in that: The following steps are involved: Step S1: Input multi-view image data and camera parameters ; In A multi-view image consists of a reference view and Source view; Step S2: In Multi-view images , a feature pyramid network FPN consisting of three stages is used to calculate the multi-scale features of multi-view images; Each stage Both use feature pyramid network FPN extraction The two-dimensional feature map of the multi-view image is obtained after the three stages are completed. ; ; Step S3: Based on the input multi-view image Generate a series of depth hypothesis planes based on the known depth range, and expand all depth hypothesis planes to the same size as the input multi-view image; Step S4: Use differentiable homography to transform multi-view image features Projected onto the expanded depth hypothesis plane, a set of reprojected features is obtained ; Based on reprojection feature Construct the cost body; Step S5: Based on the cost body obtained in step S4, the cost body is aggregated using the deformable large core attention network DLA-Net of the global information normalization module to obtain a probability body, and the probability body is normalized to obtain an initial depth map; Repeat steps S2 to S5 to obtain a final depth map; Step S6: During the unsupervised learning training process, the final depth map is optimized using photometric consistency loss, contrast consistency loss optimization, and perceptual consistency loss respectively; Step S7, repeating steps S2 to S6 until depth maps of all viewing angles are calculated and fused to output a globally consistent 3D point cloud reconstruction result.

2. The unsupervised multi-view stereo reconstruction method for anti-occlusion regions according to claim 1, characterized in that: The multi-view image data in step S1 and camera parameters The expression is as follows: ； ； In A multi-view image includes a reference view and Source View ; in, express The serial number of the multi-view image in, ; express The number of multi-view images in represents the internal parameter matrix of the kth camera, ; and Respectively represent the first camera and the The rotation matrix and translation matrix between the cameras.

3. The unsupervised multi-view stereo reconstruction method for anti-occlusion regions according to claim 1, characterized in that: The feature pyramid network FPN in step S2 introduces a channel attention module CA and a spatial attention module SA; Feature maps of multi-view images The resolution is ; in, Indicates the width of the multi-view image; Indicates the height of the multi-view image; Indicates the number of stages.

4. The unsupervised multi-view stereo reconstruction method for anti-occlusion regions according to claim 1, characterized in that: The depth assumption planes obtained in step S3 are distributed at equal intervals within the interval between the minimum depth value and the maximum depth value.

5. The unsupervised multi-view stereo reconstruction method for anti-occlusion regions according to claim 1, characterized in that: Step S4 reprojects the features With reference view Corresponding features Compare them to construct the cost body.

6. The unsupervised multi-view stereo reconstruction method for anti-occlusion regions according to claim 1, characterized in that: In the step S5, the deformable large core attention module first extracts preliminary features through 3D convolution, and passes through an activation function to enhance nonlinear characteristics, then captures spatial offset features through two layers of deformable deep convolution, then uses a layer of 3D convolution to fuse the features of multiple branches, and uses residual connection; finally, another layer of 3D convolution is used to further fuse the features; The specific contents are as follows: Step S5.1: First, use 3D convolution to extract features from the cost volume obtained in step S4. Then, use 3D deconvolution to upsample the feature map. Then, combine the upsampling with the deformable large-core attention module in step S5.1, and use deconvolution again to gradually restore the spatial resolution of the feature map. After the above steps, the regularized cost volume is obtained. ; Finally, after a layer The convolution operation of the deformable large core attention network outputs the normalized probability volume , the probability volume represents the matching degree of pixels under different depth assumptions; the normalization formula is: ； in, Depth Hypothesis Plane The probability body under represents the number of depth hypothesis planes; Step S5.2: Probability body Perform depth regression to obtain a depth map : 。 7. The unsupervised multi-view stereo reconstruction method for anti-occlusion regions according to claim 1, characterized in that: The photometric consistency loss function expression in step S6 is as follows: ； in, represents the reference view based on the input in step S1, Indicates Unwarped image.

8. The unsupervised multi-view stereo reconstruction method for anti-occlusion regions according to claim 1, characterized in that: The step S6 constructs an unsupervised comparison consistency loss function and provides a pseudo-supervisory signal in detail as follows: First, a center point is randomly selected in the image and its size is set. A connected region is generated based on the connected component analysis algorithm, and the connected region does not exceed the image boundary. Then, the connected region generated by the dilation operation is applied to the initial occlusion mask, and the pixels in the region are marked as 0, indicating that they are occluded, and the number of occluded pixels is dynamically updated. Finally, the connected region is marked as the occluded region, and the rest is marked as the unoccluded region, thereby generating a block-structured occlusion mask. Multiply the multi-view images by the corresponding structured occlusion masks element-wise to generate a set of enhanced samples ； The multi-view image data of step S1 As a regular sample, combined with enhanced samples , both are processed by steps S2 to S5 to obtain the depth map of the conventional sample And the depth map of the enhanced sample ; Finally, the contrast consistency loss between the two types of depth maps is calculated , providing a pseudo-supervisory signal for the depth estimation process: ； Represents a binary significance mask.

9. The unsupervised multi-view stereo reconstruction method for anti-occlusion regions according to claim 1, characterized in that: The detailed steps of constructing the perceptual consistency loss function in step S6 are as follows: First, use the pre-trained feature extraction network VGG19 to extract the reference view High-level semantic features ; Fine-tune the feature extraction network VGG19, that is, add a discard layer at the end of the feature extraction network VGG19 to extract the inverse distorted image obtained in step S6 High-level semantic features ; in, Indicates The high-level semantic features of the dewarped image; the dewarped image refers to the intermediate image generated when the final depth map is optimized using the photometric consistency loss function; Use perceptual consistency loss to enforce feature semantic similarity between the reference image and the unwarped image , providing effective pseudo-supervision signals for model learning: ； in, The weight representing the perceptual consistency loss.

Images