Deep learning semantic segmentation model training method applied to phase unwrapping

Abstract

The invention provides a training method of a deep learning semantic segmentation model applied to phase unwrapping, the deep learning semantic segmentation model comprises a MobileViT network, an edge information enhancement network and a cavity pyramid segmentation network, and the training method comprises the following steps: performing gamma correction on a phase shift grating image collected by a camera, a corresponding wrapped phase diagram is obtained; for each wrapped phase diagram, utilizing an adjacent pixel difference method to obtain a stripe series diagram corresponding to the wrapped phase diagram, and constructing a training image set; and sequentially inputting the wrapped phase diagrams in the training image set into the deep learning semantic segmentation model to perform feature extraction, taking the stripe series diagram in the training image set as output, and adjusting parameters of the deep learning semantic segmentation model to obtain an optimal deep learning semantic segmentation model. The model obtained through the method can improve the prediction precision and noise robustness of the fringe order, and reduces the calculation complexity and the time delay at the same time.

Description

Technical Field

[0001] The present application belongs to the field of three-dimensional reconstruction technology, and in particular, relates to a training method for a deep learning semantic segmentation model applied to phase unfolding. Background Art

[0002] Structured light 3D reconstruction technology is a non-contact measurement method that uses projected light and a camera to obtain 3D information of an object. This method has the advantages of fast speed, high accuracy, and easy operation, so it has been widely used in industrial inspection, cultural relics protection, medical imaging and other fields.

[0003] Phase unwrapping is a key step in structured light 3D reconstruction. It is mainly used to recover the true phase value from the phase wrapped phase map. The key point from the wrapped phase to the unfolded phase is the fringe order. Its accuracy directly affects the quality of the 3D model. However, traditional phase estimation algorithms are very sensitive to noise, especially in high-frequency phase information and complex scenes. Noise can cause error accumulation during phase unwrapping and phase unwrapping. Moreover, for highly discontinuous phase maps and nonlinear phase changes, the absolute phase cannot be accurately obtained, thus affecting the accuracy of 3D reconstruction.

[0004] With the rapid development of deep learning technology, neural network-based image processing methods have shown excellent performance in many computer vision tasks. Applying deep learning technology to phase unwrapping has become an important research direction to improve the accuracy and efficiency of 3D reconstruction. However, in the current research on structured light phase unwrapping, there is still a lack of an effective method that can improve the robustness and real-time performance of the algorithm while ensuring high-precision phase unwrapping. Summary of the invention

[0005] In view of this, the purpose of this application is to provide a training method for a deep learning semantic segmentation model for phase unfolding, which eliminates the error caused by the nonlinear effect of the projector by introducing gamma correction, and uses the MobileViT network, edge information enhancement network and hole pyramid segmentation network, combined with deep learning technology, to construct a deep learning semantic segmentation model that wraps the mapping relationship between phase and fringe level. This model can improve the prediction accuracy and noise robustness of the fringe level, while reducing computational complexity and time delay, and enhancing the real-time performance and adaptability of the algorithm, thereby providing a reliable technical guarantee for three-dimensional reconstruction applications with high precision and real-time requirements.

[0006] The present application provides a training method for a deep learning semantic segmentation model applied to phase unwrapping, wherein the deep learning semantic segmentation model includes: a MobileViT network, an edge information enhancement network, and a hole pyramid segmentation network, and the training method includes:

[0007] Perform gamma correction on the phase shift grating image captured by the camera to obtain the corresponding wrapped phase image;

[0008] For each wrapped phase image, the adjacent pixel difference method is used to obtain the fringe level image corresponding to the wrapped phase image, and a training image set is constructed;

[0009] The wrapped phase images in the training image set are sequentially input into the deep learning semantic segmentation model for feature extraction, and the fringe level image in the training image set is used as output, and the parameters of the deep learning semantic segmentation model are adjusted to obtain the optimal deep learning semantic segmentation model.

[0010] Furthermore, the gamma correction is performed on the phase shift grating image captured by the camera to obtain a corresponding wrapped phase image, including:

[0011] Reading pixel values ​​of the phase-shift grating image; the pixel values ​​are composed of brightness values ​​corresponding to the red, green, and blue color channels, respectively, and the brightness value has a value range of (0, 255);

[0012] Based on the pixel values ​​of the phase-shift grating image, a nonlinear transformation is performed using a gamma correction formula to eliminate errors caused by a nonlinear effect of the projector, thereby obtaining a corrected pixel value;

[0013] Based on the corrected pixel values, the wrapped phase information is calculated to obtain a corresponding wrapped phase map.

[0014] Furthermore, the step of sequentially inputting the wrapped phase images in the training image set into the deep learning semantic segmentation model for feature extraction includes:

[0015] A feature extraction strategy is constructed based on the MobileViT network and the edge information enhancement network, and feature extraction and fusion are performed on the wrapped phase image by downsampling to obtain a multi-scale wrapped phase feature image;

[0016] The atrous pyramid segmentation network is used to extract and concatenate multi-scale wrapped phase feature maps through upsampling to obtain a comprehensive wrapped phase feature map containing multi-scale wrapped phase information.

[0017] Furthermore, the feature extraction strategy is constructed based on the MobileViT network and the edge information enhancement network, and the feature extraction and fusion of the wrapped phase map are performed by downsampling to obtain a multi-scale wrapped phase feature map, including:

[0018] Downsampling the wrapped phase image to obtain initial feature images of the wrapped phase image at different scales and corresponding absolute position information; the downsampling ratios are: 1 / 2, 1 / 4, 1 / 8, 1 / 16, 1 / 32;

[0019] For the initial feature map at each scale, the MobileViT network is used to fuse the global features and local features of the initial feature map, and redundancy removal and noise suppression are performed to obtain a fused feature map.

[0020] For the fused feature map at each scale, the edge information enhancement network is used to give the fused feature map edge features to obtain a wrapped phase feature map containing image edge features.

[0021] Furthermore, the MobileViT network includes:

[0022] An n×n convolution layer and a 1×1 convolution layer, used for performing local feature representation on the initial feature map;

[0023] A global feature representation module, used for performing global feature representation based on the local features in combination with a lightweight Transformer;

[0024] The 1×1 convolution layer and the n×n convolution layer are used to fuse the local features and the global features to obtain a fused feature map.

[0025] Furthermore, the edge information enhancement network includes:

[0026] An edge feature fusion module, used for multiplying the extracted original image feature and edge feature of the fusion feature map, adding the extracted original image feature to the original image feature, and performing a 3×3 convolution operation to obtain an initial edge fusion feature;

[0027] A weight extraction module, used to extract the attention weight of the initial edge fusion feature;

[0028] The weight fusion module is used to multiply the initial edge fusion feature and the corresponding attention weight, and perform a 1×1 convolution operation to obtain a wrapped phase feature map containing image edge features.

[0029] Furthermore, the method uses a dilated pyramid segmentation network and performs feature extraction and splicing on multi-scale wrapped phase feature maps by upsampling to obtain a wrapped phase comprehensive feature map containing multi-scale wrapped phase information, including:

[0030] Using the dilated pyramid segmentation network, the features of the multi-scale wrapped phase feature map are extracted to obtain channel fusion feature maps with the same scale under each parallel channel;

[0031] Through upsampling, the channel fusion feature maps with the same scale are connected channel by channel to integrate the features of each parallel channel to obtain the wrapped phase comprehensive feature map.

[0032] Furthermore, the hollow pyramid segmentation network includes:

[0033] The parallel 1×1 convolutional layer, 3×3 convolutional layer with dilation rates of 6, 12, and 18, and pooling layer are used to extract the multi-scale wrapped phase feature map and obtain the first feature map with the same scale under each parallel channel;

[0034] The global average pooling layer and the fully connected layer are used to calculate the average value of the first feature map of each parallel channel and then perform a splicing operation to obtain a second feature map;

[0035] The channel mixing layer, the attention layer and the fully connected layer are used to mix the second feature maps of each parallel channel, and after assigning weights to each parallel channel through the attention layer, fuse them with the first feature map to obtain channel fusion feature maps with the same scale under each parallel channel.

[0036] Furthermore, the step of taking the fringe level map in the training image set as output and adjusting the parameters of the deep learning semantic segmentation model to obtain an optimal deep learning semantic segmentation model includes:

[0037] The fringe level map in the training image set is taken as output, and the parameters of the deep learning semantic segmentation model are adjusted based on the wrapped phase comprehensive feature map to obtain the optimal deep learning semantic segmentation model.

[0038] The training method of the deep learning semantic segmentation model for phase unwrapping provided in the present application eliminates the error caused by the nonlinear effect of the projector by introducing gamma correction, and at the same time utilizes the MobileViT network, the edge information enhancement network and the hole pyramid segmentation network, combined with deep learning technology, to construct a deep learning semantic segmentation model that wraps the mapping relationship between the phase and the fringe level. The model can improve the prediction accuracy and noise robustness of the fringe level, while reducing the computational complexity and time delay, and enhancing the real-time performance and adaptability of the algorithm, thereby providing a reliable technical guarantee for 3D reconstruction applications with high precision and high real-time requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A flowchart of a training method for a deep learning semantic segmentation model applied to phase unwrapping provided in an embodiment of the present application is shown;

[0040] Figure 2A schematic diagram of the structure of the MobileViT network provided in an embodiment of the present application is shown;

[0041] Figure 3 A schematic diagram of the structure of an edge information enhancement network provided in an embodiment of the present application is shown;

[0042] Figure 4 A schematic diagram of the structure of a hollow pyramid segmentation network provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the technical solution more clear, the technical solution is further described in detail below in conjunction with specific implementation methods. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the technical solution.

[0044] In the structured light 3D reconstruction technology, the relationship between the unwrapped phase, the wrapped phase and the fringe level in the phase unwrapping process satisfies the following formula (1):

[0045]

[0046] Where φ(x,y) is the unwrapped phase, is the wrapping phase, and k(x,y) is the fringe order.

[0047] It can be seen from formula (1) that the key point from wrapped phase to unwrapped phase is the fringe order. Therefore, high-precision determination of the fringe order is of great significance to achieve phase unwrapping.

[0048] In view of this, the present application embodiment provides a training method for a deep learning semantic segmentation model applied to phase unwrapping to establish a mapping relationship between the wrapped phase and the fringe level. Figure 1 , Figure 1 FIG. 2 is a flow chart of a method for training a deep learning semantic segmentation model for phase unwrapping provided in an embodiment of the present application. Figure 1 As shown, the deep learning semantic segmentation model includes: MobileViT network, edge information enhancement network and hollow pyramid segmentation network, and the training method includes:

[0049] S101, performing gamma correction on the phase shift grating image acquired by the camera to obtain a corresponding wrapped phase image.

[0050] In specific implementation, the corresponding wrapped phase diagram is obtained through the following steps:

[0051] Step 1011, read the pixel value of the phase-shift grating image; the pixel value is composed of brightness values ​​corresponding to the red, green and blue color channels respectively, and the value range of the brightness value is (0, 255).

[0052] In this step, the pixel values ​​of the phase-shift grating images taken by multiple cameras are read. Here, the phase-shift grating image is usually stored in RGB format, and each pixel value is composed of brightness values ​​corresponding to the red, green, and blue color channels respectively, and the brightness value range is (0, 255).

[0053] Step 1012: Based on the pixel values ​​of the phase-shift grating image, a gamma correction formula is used to perform a nonlinear transformation to eliminate errors caused by the nonlinear effect of the projector, thereby obtaining corrected pixel values.

[0054] In this step, a suitable gamma value is selected according to the device characteristics of the camera. Here, the gamma value range is (0.8, 2.5). For each brightness value that constitutes the pixel value, the following gamma correction formula (2) is used to obtain the brightness value of the corrected pixel value:

[0055] I out =(I in / 255) γ ×255; (2)

[0056] In the formula, I in is the pixel value of the phase shift grating image, γ is the gamma value, I out is the pixel value after correction.

[0057] Step 1013: Calculate the wrapped phase information based on the corrected pixel value to obtain a corresponding wrapped phase map.

[0058] In this step, the phase information is calculated using the following formula (3) to obtain the corresponding wrapped phase map:

[0059]

[0060] In the formula, is the wrapping phase, I n (x, y) is the grayscale value of the nth phase-shift fringe pattern, and N is the number of phase-shift fringe patterns.

[0061] S102 , for each wrapped phase image, using an adjacent pixel difference method, obtain a fringe level image corresponding to the wrapped phase image, and construct a training image set.

[0062] In this step, the phase difference between adjacent pixels of the wrapped phase image is calculated using the adjacent pixel difference method. If the phase difference is not within the range of (-π, π), it is considered that a jump has occurred. Then, the fringe order corresponding to the wrapped phase image is determined according to the positive and negative and magnitude of the phase difference, and a fringe order map is obtained. Further, the one-to-one corresponding wrapped phase image and fringe order map are divided into a training set, a validation set, and a test set in a ratio of 8:1:1. Here, the division requirements are: the number of training sets is not less than 120 pairs of sample images, the validation set is not less than 15 pairs of sample images, and the test set is not less than 15 pairs of sample images.

[0063] S103, inputting the wrapped phase map in the training image set into the deep learning semantic segmentation model in sequence to perform feature extraction, and taking the fringe level map in the training image set as output, adjusting the parameters of the deep learning semantic segmentation model to obtain the optimal deep learning semantic segmentation model.

[0064] In this step, the deep learning semantic segmentation model provided in the embodiment of the present application is divided into an encoding stage and a decoding stage. The encoding stage generates a high-level semantic representation based on the wrapped phase map using the MobileVIT network and the edge information enhancement network. The decoding stage uses the dilated pyramid segmentation network to extract finer-grained multi-scale spatial information from the semantic representation obtained in the encoding stage and restore its resolution to output a fringe level map.

[0065] In a specific implementation, the wrapped phase images in the training image set are sequentially input into the deep learning semantic segmentation model through the following steps to perform feature extraction:

[0066] Step 1031: construct a feature extraction strategy based on the MobileViT network and the edge information enhancement network, and perform feature extraction and fusion on the wrapped phase map by downsampling to obtain a multi-scale wrapped phase feature map.

[0067] In the specific implementation, the multi-scale wrapped phase feature map is obtained through the following steps:

[0068] Step 201: downsample the wrapped phase image to obtain initial feature images of the wrapped phase image at different scales and corresponding absolute position information.

[0069] The downsampling ratios are: 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 respectively.

[0070] Step 202: For each initial feature map at each scale, the global features and local features of the initial feature map are fused using the MobileViT network, and redundancy removal and noise suppression are performed to obtain a fused feature map.

[0071] Wherein, the MobileViT network includes:

[0072] An n×n convolution layer and a 1×1 convolution layer, used for performing local feature representation on the initial feature map;

[0073] A global feature representation module, used for performing global feature representation based on the local features in combination with a lightweight Transformer;

[0074] The 1×1 convolution layer and the n×n convolution layer are used to fuse the local features and the global features to obtain a fused feature map.

[0075] In this step, see Figure 2 , Figure 2 The figure shows a schematic diagram of the structure of the MobileViT network provided in the embodiment of the present application. Figure 2 As shown in the figure, the workflow of the MobileViT network is as follows: 1) For the initial feature map x∈R at each scale H ×W×C , through n×n convolution layer and 1×1 convolution layer, local feature representation is performed. 2) Global feature representation module, firstly represents the local feature X∈R H×W×C , reshaped into a flattened patch sequence X f ∈R N×PC , and project it into a fixed d-dimensional space to obtain X p ∈R N×d ; Then use the lightweight Transformer as convolution to learn the global representation; Finally, compress the patch sequence, divide the feature map into blocks of fixed size, and add the pixel values ​​in each block to obtain a single value; where H, W, C represent the height, width and number of channels of the initial feature map, P = wh is the number of pixels in the patch, N is the number of patches, and the computational cost of self-attention is O(N 2 d). 3) Through the 1×1 convolution layer and the n×n convolution layer, the global features and local features are fused to obtain the fused feature map.

[0076] Step 203: for each fused feature map at each scale, an edge information enhancement network is used to give edge features to the fused feature map, so as to obtain a wrapped phase feature map containing image edge features.

[0077] Wherein, the edge information enhancement network includes:

[0078] An edge feature fusion module, used for multiplying the extracted original image feature and edge feature of the fusion feature map, adding the extracted original image feature to the original image feature, and performing a 3×3 convolution operation to obtain an initial edge fusion feature;

[0079] A weight extraction module, used to extract the attention weight of the initial edge fusion feature;

[0080] The weight fusion module is used to multiply the initial edge fusion feature and the corresponding attention weight, and perform a 1×1 convolution operation to obtain a wrapped phase feature map containing image edge features.

[0081] In this step, see Figure 3 , Figure 3 FIG. 1 is a schematic diagram of the structure of the edge information enhancement network provided by the embodiment of the present application. Figure 3 As shown in the figure, the workflow of the edge information enhancement network is as follows: 1) extract the original image features f of the fused feature map x and edge features f e , after multiplying the two features, use a skip connection to combine them with the original image feature f x 1) The initial edge fusion feature is firstly subjected to a global average pooling operation to reduce the impact of overfitting, and then subjected to a 1×1 convolution operation and a Sigmod activation function to obtain the attention weight of the initial edge fusion feature. 2) The initial edge fusion feature is first subjected to a global average pooling operation to reduce the impact of overfitting, and then subjected to a 1×1 convolution operation and a Sigmod activation function to obtain the attention weight of the initial edge fusion feature. 3) The initial edge fusion feature is multiplied by the corresponding attention weight, and then subjected to a 1×1 convolution operation to obtain the wrapped phase feature f containing the image edge feature. y , and the corresponding wrapped phase feature map.

[0082] Step 1032: Use the dilated pyramid segmentation network and up-sample to extract and concatenate the multi-scale wrapped phase feature maps to obtain a wrapped phase comprehensive feature map containing multi-scale wrapped phase information.

[0083] In specific implementation, the following steps are performed to obtain a comprehensive wrapped phase feature map containing multi-scale wrapped phase information:

[0084] Step 301: using the dilated pyramid segmentation network, extracting features of multi-scale wrapped phase feature maps, and obtaining channel fusion feature maps with the same scale under each parallel channel.

[0085] Wherein, the hollow pyramid segmentation network includes:

[0086] The parallel 1×1 convolutional layer, 3×3 convolutional layer with dilation rates of 6, 12, and 18, and pooling layer are used to extract the multi-scale wrapped phase feature map and obtain the first feature map with the same scale under each parallel channel;

[0087] The global average pooling layer and the fully connected layer are used to calculate the average value of the first feature map of each parallel channel and then perform a splicing operation to obtain a second feature map;

[0088] The channel mixing layer, the attention layer and the fully connected layer are used to perform a mixing operation on the second feature maps of each parallel channel, and after assigning weights to each parallel channel through the attention layer, fuse them with the first feature map to obtain a channel fusion feature map with the same scale under each parallel channel.

[0089] In this step, see Figure 4 , Figure 4 The structure diagram of the hollow pyramid segmentation network provided by the embodiment of the present application is shown in FIG. Figure 4 As shown in the figure, the workflow of the dilated pyramid segmentation network is as follows: 1) Use parallel 1×1 convolution layers, 3×3 convolution layers with dilation rates of 6, 12, and 18, and pooling layers to extract the features of the wrapped phase feature maps of multiple scales (i.e., different resolutions), so that the model can focus on local details and global information, and obtain the first feature maps with the same scale under each parallel channel. 2) After calculating the average value of each first feature map using the global average pooling layer, the obtained two-dimensional feature maps are spliced ​​to pass the information to the next processing stage and obtain the second feature map. 3) Use the channel mixing layer to mix the second feature maps of each parallel channel; use the attention layer to learn and assign weights to each parallel channel in the mixed channel features, so that the network can focus on more important channel information; use the fully connected layer to fuse the channel features with the first feature map to obtain a channel fusion feature map.

[0090] Step 302: connect the channel fusion feature maps with the same scale by upsampling to integrate the features of each parallel channel and obtain a wrapped phase comprehensive feature map.

[0091] In this step, the channel fusion feature maps adjusted to the same scale are connected through upsampling to integrate the feature information of different stages and obtain the wrapped phase comprehensive feature map.

[0092] In the specific implementation, the optimal deep learning semantic segmentation model is obtained through the following steps:

[0093] The fringe level map in the training image set is taken as output, and the parameters of the deep learning semantic segmentation model are adjusted based on the wrapped phase comprehensive feature map to obtain the optimal deep learning semantic segmentation model.

[0094] In this step, the fringe level map in the training image set is used as the output, and the relationship between the wrapped phase comprehensive feature mapping and the fringe level is learned through the convolution layer to obtain the optimal deep learning semantic segmentation model. The deep learning semantic segmentation model can output the fringe level map based on the input wrapped phase map.

[0095] The above contents are only preferred embodiments of the present invention. For ordinary technicians in this field, many changes can be made in the specific implementation methods and application scopes based on the ideas of the present technical content. As long as these changes do not deviate from the concept of the present invention, they all fall within the scope of protection of this patent.

Claims

1. A training method for a deep learning semantic segmentation model applied to phase unwrapping, characterized in that: The deep learning semantic segmentation model includes: a MobileViT network, an edge information enhancement network and a hole pyramid segmentation network, and the training method includes: Perform gamma correction on the phase shift grating image captured by the camera to obtain the corresponding wrapped phase image; For each wrapped phase image, the adjacent pixel difference method is used to obtain the fringe level image corresponding to the wrapped phase image, and a training image set is constructed; The wrapped phase images in the training image set are sequentially input into the deep learning semantic segmentation model for feature extraction, and the fringe level image in the training image set is used as output, and the parameters of the deep learning semantic segmentation model are adjusted to obtain the optimal deep learning semantic segmentation model.

2. The method according to claim 1, characterized in that The gamma correction is performed on the phase shift grating image acquired by the camera to obtain a corresponding wrapped phase image, including: Reading pixel values ​​of the phase-shift grating image; the pixel values ​​are composed of brightness values ​​corresponding to the red, green, and blue color channels, respectively, and the brightness value has a value range of (0, 255); Based on the pixel values ​​of the phase-shift grating image, a nonlinear transformation is performed using a gamma correction formula to eliminate errors caused by a nonlinear effect of the projector, thereby obtaining a corrected pixel value; Based on the corrected pixel values, the wrapped phase information is calculated to obtain a corresponding wrapped phase map.

3. The method according to claim 1, characterized in that The step of sequentially inputting the wrapped phase images in the training image set into the deep learning semantic segmentation model to extract features includes: A feature extraction strategy is constructed based on the MobileViT network and the edge information enhancement network, and feature extraction and fusion are performed on the wrapped phase image by downsampling to obtain a multi-scale wrapped phase feature image; The atrous pyramid segmentation network is used to extract and concatenate multi-scale wrapped phase feature maps through upsampling to obtain a comprehensive wrapped phase feature map containing multi-scale wrapped phase information.

4. The method according to claim 3, characterized in that The feature extraction strategy is constructed based on the MobileViT network and the edge information enhancement network, and the feature extraction and fusion of the wrapped phase map are performed by downsampling to obtain a multi-scale wrapped phase feature map, including: Downsampling the wrapped phase image to obtain initial feature images of the wrapped phase image at different scales and corresponding absolute position information; the downsampling ratios are: 1 / 2, 1 / 4, 1 / 8, 1 / 16, 1 / 32; For the initial feature map at each scale, the MobileViT network is used to fuse the global features and local features of the initial feature map, and redundancy removal and noise suppression are performed to obtain a fused feature map. For the fused feature map at each scale, the edge information enhancement network is used to give the fused feature map edge features to obtain a wrapped phase feature map containing image edge features.

5. The method according to claim 4, characterized in that The MobileViT network includes: An n×n convolution layer and a 1×1 convolution layer, used for performing local feature representation on the initial feature map; A global feature representation module, used for performing global feature representation based on the local features in combination with a lightweight Transformer; The 1×1 convolution layer and the n×n convolution layer are used to fuse the local features and the global features to obtain a fused feature map.

6. The method according to claim 4, characterized in that The edge information enhancement network comprises: An edge feature fusion module, used for multiplying the extracted original image feature and edge feature of the fusion feature map, adding the extracted original image feature to the original image feature, and performing a 3×3 convolution operation to obtain an initial edge fusion feature; A weight extraction module, used to extract the attention weight of the initial edge fusion feature; The weight fusion module is used to multiply the initial edge fusion feature and the corresponding attention weight, and perform a 1×1 convolution operation to obtain a wrapped phase feature map containing image edge features.

7. The method according to claim 3, characterized in that The method uses a dilated pyramid segmentation network and performs feature extraction and splicing on multi-scale wrapped phase feature maps by upsampling to obtain a wrapped phase comprehensive feature map containing multi-scale wrapped phase information, including: Using the dilated pyramid segmentation network, the features of the multi-scale wrapped phase feature map are extracted to obtain channel fusion feature maps with the same scale under each parallel channel; Through upsampling, the channel fusion feature maps with the same scale are connected channel by channel to integrate the features of each parallel channel to obtain the wrapped phase comprehensive feature map.

8. The method according to claim 7, characterized in that The hollow pyramid segmentation network includes: The parallel 1×1 convolutional layer, 3×3 convolutional layer with dilation rates of 6, 12, and 18, and pooling layer are used to extract the multi-scale wrapped phase feature map and obtain the first feature map with the same scale under each parallel channel; The global average pooling layer and the fully connected layer are used to calculate the average value of the first feature map of each parallel channel and then perform a splicing operation to obtain a second feature map; The channel mixing layer, the attention layer and the fully connected layer are used to mix the second feature maps of each parallel channel, and after assigning weights to each parallel channel through the attention layer, fuse them with the first feature map to obtain channel fusion feature maps with the same scale under each parallel channel.

9. The method according to claim 7, characterized in that The step of taking the fringe level map in the training image set as output and adjusting the parameters of the deep learning semantic segmentation model to obtain an optimal deep learning semantic segmentation model comprises: The fringe level map in the training image set is taken as output, and the parameters of the deep learning semantic segmentation model are adjusted based on the wrapped phase comprehensive feature map to obtain the optimal deep learning semantic segmentation model.

Images