A deep learning-based single-frame digital fringe end-to-end phase unwrapping method

Abstract

The application provides a kind of single frame digital fringe end-to-end phase analysis method based on deep learning, comprising the following steps: first, build a digital twin based fringe projection system, construct the deformed fringe-absolute phase dataset;Then construct IncepC-Unet network model, the deformed fringe-absolute phase dataset is proportioned into test set and training set, train and test IncepC-Unet network model;Finally, the single frame deformed fringe image of the measured object is predicted by the above trained IncepC-Unet network model.The single frame digital fringe end-to-end phase analysis method designed by the application completely abandons the traditional cascade type analysis process of absolute phase unwrapping, realizes the absolute phase calculation of single frame digital fringe end-to-end, improves the efficiency and accuracy of phase analysis, and promotes the development of three-dimensional measurement towards intelligent direction.

Description

Technical Field

[0001] This invention belongs to the field of optical three-dimensional measurement technology, specifically relating to a single-frame digital stripe end-to-end phase analysis method based on deep learning. Background Technology

[0002] Currently, 3D measurement technology plays a vital role in various fields, and it is widely used in aerospace manufacturing, smart healthcare, intelligent recognition, and artificial intelligence. Fringe projection 3D imaging, due to its advantages such as non-contact operation, high precision, full-field measurement, and high point cloud reconstruction efficiency, has become the mainstream optical method in 3D measurement technology. However, traditional fringe projection 3D imaging obtains absolute phase information by projecting multiple sets of grating fringes and combining phase-shifting / time-phase unrolling methods. This requires fringe image acquisition, phase demodulation, and phase resolution, which inevitably prolongs data acquisition time and significantly impacts the efficiency of 3D information acquisition. In contrast, spatial domain methods only require one frame of fringes and its corresponding principal phase to recover the absolute phase. Early algorithms for extracting the principal phase include Fourier transform (FT) and windowed Fourier transform (WFT); methods for obtaining the absolute phase mainly include branching, quality map-guided methods, and unweighted FFT. Although there are many mature phase resolution methods based on the spatial domain method, the phase resolution is affected by the influence of adjacent pixels, which leads to the gradual propagation of errors. When facing targets with complex structures, discontinuous or isolated surface characteristics, it is easy to cause large phase resolution errors. The absolute phase is often accompanied by serious phenomena such as "stringing" and "packet loss". The reconstruction accuracy is still far inferior to the solution method of multiple fringe projection.

[0003] Therefore, how to quickly, accurately, and unambiguously obtain the absolute phase of a target using single-frame fringe images and faster methods is an urgent problem to be solved in the field of fringe projection 3D imaging.

[0004] In recent years, with the development of artificial intelligence technology, phase analysis methods based on deep learning technology have achieved significant breakthroughs. In phase demodulation, a single-frame fringe analysis method based on deep learning utilizes two convolutional neural networks (CNN1 and CNN2) to learn and generate the sine (M) and cosine (D) components required for phase demodulation from the acquired raw fringe image. These sine and cosine results are then substituted into the arctangent function to calculate the wrapped phase. Based on this, multi-path CNN+ResNet networks have been successively used to predict the M / D ratio of the wrapped phase. Compared with traditional FT and WFT, this method can extract phase information more accurately using only one fringe image, especially for object surfaces with rich details. In phase unwrapping, there are mainly indirect unwrapping methods that predict the fringe order and direct unwrapping methods that predict the absolute phase. The former learns from one or more frames of wrapped phase, first inferring or segmenting the fringe order image required for the absolute phase, and then combining it with the wrapped phase to recover the absolute phase information; the latter learns from one frame of wrapped phase and directly generates the absolute phase information.

[0005] Currently, in phase resolution, most deep learning methods only focus on phase demodulation or phase unrolling at a certain stage of phase resolution, or fuse the two using a multi-stage network. Such a measurement process is relatively complex and inefficient, and also suffers from the problem of error forward propagation. Summary of the Invention

[0006] The purpose of this invention is to solve the problems of complex measurement process, low efficiency and forward propagation error in the existing technology of obtaining absolute phase using single-frame fringe images and phase analysis. Instead, it provides a deep learning-based end-to-end phase analysis method for single-frame digital fringe images.

[0007] To achieve the above objectives, the technical solution provided by this invention is:

[0008] A deep learning-based end-to-end phase analysis method for single-frame digital stripes includes the following steps:

[0009] Step 1: Build a fringe projection system based on digital twin, acquire deformed fringe images using the phase shift method, calculate the absolute phase of the deformed fringe images using the dual-frequency phase deconvolution algorithm, and construct a deformed fringe-absolute phase dataset.

[0010] Step 2: Construct the InncepC-Unet network model, including the encoder network, decoder network, and skip connections;

[0011] Step 3: Divide the deformed stripe-absolute phase dataset from Step 1 into a training set and a test set according to a certain ratio. Train the InncepC-Unet network model based on the training set and test the trained InncepC-Unet network model based on the test set. Based on the training and test results, optimize and adjust the relevant parameter settings of the InncepC-Unet network model.

[0012] Step 4: Use the InncepC-Unet network model designed in Step 3 to perform absolute phase prediction on the single-frame deformed stripe image of the object under test.

[0013] Furthermore, in step 1, the stripe projection system includes a camera, a projector, and an object to be measured, wherein the camera and the projector are placed horizontally, and the object to be measured is perpendicular to the projector.

[0014] Furthermore, in step 1, the steps for constructing the deformed fringe-absolute phase dataset include:

[0015] Step 1.1: Render the object to be tested using high-frequency and low-frequency sinusoidal fringes respectively, and then use a camera to capture and save deformed fringe images. The first frame of the high-frequency deformed fringe image is a single-frame deformed fringe image.

[0016] Step 1.2: For the high-frequency deformed stripe images saved in Step 1.1, excluding the single-frame deformed stripe images, calculate their absolute phase using the dual-frequency phase analysis algorithm and save them;

[0017] Step 1.3: Replace the object to be tested and repeat the operations of steps 1.1 to 1.2. Several single-frame deformed stripe images and absolute phases are finally combined to form a deformed stripe-absolute phase dataset.

[0018] Furthermore, in step 2, the coding network consists of five coding layers. The first four coding layers each consist of an Inception residual block, a coordinate attention block, and a 2×2 max pooling block. The fifth coding layer consists of an Inception residual block.

[0019] The decoding network consists of four decoding layers, each of which consists of a transposed convolution, a concatenation block, and an Inception residual block;

[0020] Skip connections are used to fuse features between the encoding and decoding layers.

[0021] Furthermore, the Inception residual block consists of a parallel layer structure composed of branches 0, 1, 2, 3, and 4. Branch 0 consists of convolutional blocks padded with 0 and a batch normalization layer, while branches 1 to 4 each consist of convolutional blocks padded with 1 and a batch normalization layer. The four C / 4 channel outputs of branches 1 to 4 are connected into a single C channel output with a specified dimension of 1, and the C channel output is added to the output of branch 0.

[0022] Furthermore, the Inception residual block in the encoding network enhances the multi-scale feature extraction capability of the InncepC-Unet network model through multi-branch pathways; the coordinate attention block is used to extract the spatial dependence of stripe images at different locations in the local feature map, and the coordinate attention block uses one-dimensional pooling in the X and Y directions respectively to replace global two-dimensional pooling, enabling the InncepC-Unet network model to learn the spatial distribution features of the image more accurately and improve the InncepC-Unet network model's attention to the effective target region of the image; 2×2 max pooling can halve the width and height of the extracted feature map, reduce the amount of computation, and improve the training speed of the InncepC-Unet network model.

[0023] Furthermore, in step 3, the training steps of the InncepC-Unet network model include:

[0024] Step 3.1: Normalize the deformed stripe-absolute phase dataset generated in Step 1, and scale the input and output of the InncepC-Unet network model. The input of the InncepC-Unet network model is a single frame deformed stripe image.

[0025] Step 3.2: Set the relevant parameters for training the InncepC-Unet network model;

[0026] Step 3.3: Input the predicted phase and the true phase into the loss function of the InncepC-Unet network model. Based on the result of the loss function, train the InncepC-Unet network model until the result of the loss function converges.

[0027] Furthermore, in step 3, the deformed stripe-absolute phase dataset is randomly divided into a training set and a test set in a 9:1 ratio, wherein the training set is not visible in the test set.

[0028] Furthermore, in step 4, the step of performing absolute phase prediction on a single frame of deformed stripe image using the InncepC-Unet network model includes:

[0029] Step 4.1: Input the single-frame deformed stripe image of the object under test into the InncepC-Unet network model;

[0030] Step 4.2: The input single-frame deformed stripe image is first processed by the encoding layer of the encoding network to extract local features and obtain a feature map. Then, the decoding layer of the decoding network performs layer-by-layer transposed convolution to obtain an intermediate feature map. The intermediate feature map is then concatenated with the feature map of the same size in the encoding network to obtain a concatenated feature map.

[0031] Step 4.3: Input the stitched feature map obtained in step 4.2 into the Inception residual block of the decoding network. The feature map with 1 channel is obtained through branch 0, which is the absolute phase corresponding to the deformed stripe image of the object under test.

[0032] The advantages of this invention are:

[0033] The single-frame digital stripe end-to-end phase analysis method designed in this invention completely abandons the traditional cascaded analysis process of absolute phase expansion, realizes end-to-end absolute phase calculation of single-frame digital stripes, improves the efficiency and accuracy of phase analysis, and promotes the development of three-dimensional measurement towards intelligence. Attached Figure Description

[0034] The features and advantages of the invention will become more readily apparent from the following description with reference to the accompanying drawings, which are not drawn to scale and some features are enlarged or reduced to show details of specific parts.

[0035] Figure 1 This is a flowchart of the end-to-end phase analysis method for single-frame digital stripes based on deep learning in an embodiment of the present invention;

[0036] Figure 2 This is the stripe projection system and measurement scenario in an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of the InncepC-Unet network model in an embodiment of the present invention;

[0038] Figure 4 This is a schematic diagram of the Inception residual block structure in the InncepC-Unet network model in this embodiment of the invention;

[0039] Figure 5 This is a schematic diagram of the coordinate attention block structure in the InncepC-Unet network model in this embodiment of the invention;

[0040] Figure 6 This is a comparison diagram of the binary phase error between the phase analysis method designed in this embodiment of the invention and three other phase analysis methods. Detailed Implementation

[0041] The present invention will now be described in detail with reference to the accompanying drawings and exemplary embodiments thereof. It should be noted that the following detailed description of the present invention is for illustrative purposes only and is not intended to limit the scope of the invention.

[0042] A deep learning-based end-to-end phase analysis method for single-frame digital stripes, such as... Figure 1 As shown, it includes the following steps:

[0043] Step 1: Build a fringe projection system based on digital twin, acquire deformed fringe images using the phase shift method, calculate the absolute phase of the deformed fringe images using the dual-frequency phase deconvolution algorithm, and construct a deformed fringe-absolute phase dataset.

[0044] Abundant datasets support the training of deep learning networks. However, creating actual datasets is typically a time-consuming task, requiring significant time and resources to collect and label real-world data. Therefore, this invention utilizes the 3D modeling software 3D Max to construct a virtual fringe projection system based on the actual measurement system. For example... Figure 2 As shown, it includes a camera, a projector, and the object to be measured. Specifically, the camera is a video camera, and the projector is a spotlight. The camera and the projector are placed horizontally, and the object to be measured is perpendicular to the projector.

[0045] The actual measurement system parameters referenced by the fringe projection system are shown in Table 1.

[0046] Table 1. Overview of parameter settings for the actual system projector and industrial camera.

[0047]

[0048] The steps to construct a deformable fringe-absolute phase dataset include:

[0049] Step 1.1: Place the object to be measured at the measurement position. The projector renders six-step phase-shift fringe image sequences with fringe frequencies of 19 and 24 onto the surface of the object to be measured. Then, a camera is used to capture twelve phase-shift deformed fringe images and save them in *.bmp file format. The first frame of the high-frequency deformed fringe image is a single-frame deformed fringe image. Convert the *.bmp file format of the single-frame deformed fringe image to *.mat file format.

[0050] The nth deformed stripe image captured by the camera can be represented as:

[0051]

[0052] Where n = 0, 1, 2, 3, ..., N-1, I n (x,y) represents the intensity of the deformed stripe image, A(x,y) represents the background intensity, and B(x,y) represents the modulation degree. The principal value of the phase is represented by 2πn / N, which represents the phase shift of the nth sine fringe.

[0053] Step 1.2: For the high-frequency deformable stripe images saved in Step 1.1, excluding the single-frame deformable stripe images, calculate their absolute phase using the dual-frequency phase analysis algorithm and save them in *.mat file format.

[0054] The wrapping phase of the above six-step phase-shifting fringe image is calculated using the least squares method, as shown in the following formula:

[0055]

[0056] The dual-frequency dephase algorithm is applied to the two different frequencies of the encapsulated phase, as shown in the following equation:

[0057]

[0058] The two fringe frequencies are respectively represented as f l (Low frequency) and f h (High frequency), the principal phases corresponding to the high-frequency and low-frequency fringe images are respectively and K h and K l k represents the order of the high-frequency and low-frequency grating fringes, respectively. l (x,y) and k h The set of integers (x, y), where Ω represents k. l (x,y)·f h -k h (x,y)·f l In the interval (-f) h f l The set of integers within ), round[x] represents rounding to the nearest integer to x, and s represents the set count.

[0059] Step 1.3: Replace the object to be tested and repeat the operations of steps 1.1 to 1.2. Several single-frame deformed fringe images and absolute phases are finally combined to form a deformed fringe-absolute phase dataset in *.mat format.

[0060] In the fringe projection system, 150 3D models of any category are selected as the objects to be tested, and they are rotated, scaled, and combined to generate 1120 pairs of deformable fringe-absolute phase datasets.

[0061] Step 2: Construct the InncepC-Unet network model, such as Figure 3 As shown, it includes an encoding network, a decoding network, and skip connections.

[0062] The coding network consists of five coding layers. The first four coding layers each consist of an Inception residual block, a coordinate attention block, and a 2×2 max pooling with a stride of 2 for sampling. The fifth coding layer consists of an Inception residual block.

[0063] The encoding network sequentially passes through five layers of Inception residual blocks and four layers of coordinate attention blocks. The Inception residual blocks enhance the multi-scale feature extraction capability of the InncepC-Unet network model through multi-branch paths. The coordinate attention blocks are used to extract the spatial dependencies of striped images at different locations in the local feature map. 2×2 max pooling halves the width and height of the extracted feature map, aiming to reduce computation and improve the training speed of the InncepC-Unet network model. Traditional global pooling operations aggregate the entire feature map; for example, average pooling averages the feature values, and max pooling only retains the largest feature value. This operation loses some local feature information within a spatial range. In phase prediction tasks, not only is the overall image information important, but also the continuity and monotonicity of local phases. Therefore, to preserve the positional information in the feature map, the coordinate attention blocks designed in this invention perform one-dimensional pooling in the X and Y directions respectively, replacing the traditional global two-dimensional pooling. This helps the InncepC-Unet network model learn the spatial distribution features of the image more accurately and improves the InncepC-Unet network model's focus on the effective target region of the image.

[0064] The decoding network consists of four decoding layers. Each decoding layer consists of a 3×3 transposed convolution with a stride of 2, a concatenation block, and an Inception residual block. The 3×3 transposed convolution is used to upsample the input feature map of the previous layer, recovering the width and height of the feature map from bottom to top.

[0065] like Figure 4 As shown, the Inception residual block consists of a parallel layer structure comprising branches 0, 1, 2, 3, and 4. Branch 0 consists of a 1×1 convolutional block padded with zeros and a batch normalization layer. Branches 1 through 4 each consist of a 3×3 convolutional block padded with one and a batch normalization layer. The four C / 4 channel outputs of branches 1 through 4 are concatenated into a single C channel output with a specified dimension of 1, and this C channel output is added to the output of branch 0. The Inception residual block extracts feature information at different scales from four convolutional layers of different depths and performs feature fusion through concatenation, which helps improve the multi-scale feature extraction of the InncepC-Unet network model and enhances its expressive power.

[0066] The skip connection after the Inception residual block of the encoding layer is used to concatenate the features of the encoding and decoding layers. The concatenated feature map is then fed back into the Inception residual block to process the high-dimensional feature map.

[0067] Step 3: Divide the deformed stripe-absolute phase dataset from Step 1 into a training set and a test set according to a certain ratio. Train the InncepC-Unet network model based on the training set, and test the trained InncepC-Unet network model based on the test set. Based on the training and test results, optimize and adjust the relevant parameter settings of the InncepC-Unet network model.

[0068] The deformable stripe-absolute phase dataset combines 145 pairs of real data and randomly divides them into a training set and a test set in a 9:1 ratio. The training set is not visible in the test set.

[0069] The InncepC-Unet network model was trained using the training set described above. The virtual system runtime environment was set up as shown in Table 2.

[0070] Table 2 Virtual System Operating Environment

[0071]

[0072] Specifically, the training steps for the InncepC-Unet network model include:

[0073] Step 3.1: Normalize the deformed stripe-absolute phase dataset generated in Step 1, scale the input and output of the InncepC-Unet network model, and use a single frame deformed stripe image as input and absolute phase as label;

[0074] Step 3.2: Set the optimizer and parameters such as learning rate, batch size, and number of iterations for training the InncepC-Unet network model, as shown in Table 3;

[0075] Table 3. Overview of training options and parameters.

[0076] 300 8 0.01 0.0001 0.9

[0077] Step 3.3: Input the network's predicted phase and true phase into the loss function of the InncepC-Unet network model. Based on the result of the loss function, train the InncepC-Unet network model until the result of the loss function converges. If the MAE function is used as the loss function, the training cycle is 250 rounds when the result of the loss function converges.

[0078] The trained InncepC-Unet network model was tested using a set of David sculptures in the test set. A comparative experiment was conducted, comparing the error of the phase analysis method designed in this invention with that of traditional Unet, U2-net, Res-Unet, and InncepC-Unet. Figure 6 As shown, Figure 6 (a) is a deformed stripe. Figure 6 (b)~ Figure 6 (e) shows the phase prediction results of the four algorithms: traditional Unet, U2-net, Res-Unet, and the algorithm of this invention. Figure 6 (f) represents the true phase distribution. Figure 6 (g)~ Figure 6 (j) shows the binary phase error distribution between the traditional Unet, U2-net, Res-Unet and the present invention and the actual phase distribution. The more black pixels there are, the greater the error; conversely, the fewer the black pixels, the smaller the error. As can be seen from the figure, the phase analysis method designed in this invention has small error and superior performance.

[0079] Step 4: Use the InncepC-Unet network model designed in Step 3 to perform absolute phase prediction on the single-frame deformed stripe image of the object under test.

[0080] Specifically, the steps for absolute phase prediction of a single-frame deformable stripe image using the InncepC-Unet network model include:

[0081] Step 4.1: Input the single-frame deformed stripe image of the object under test into the InncepC-Unet network model;

[0082] Step 4.2: The input single-frame deformed stripe image is first processed through five coding layers of the coding network to extract local features and obtain a feature map. Specifically: (e.g., ...) Figure 5 As shown, the coordinate attention block first performs one-dimensional pooling on the X and Y directions of the extracted feature map, then concatenates them. Next, it uses a 1×1 convolutional block and a non-linear activation function from the Inception residual block to obtain a transitional feature map. Finally, it multiplies the feature maps segmented in the X and Y directions with the original input single-frame deformed stripe image to obtain the final output feature map of the coordinate attention block. Then, the decoding layer of the decoding network performs layer-by-layer transposed convolutions to obtain intermediate feature maps. These intermediate feature maps are then concatenated with feature maps of the same size from the encoding network via skip connections after the Inception residual block of the encoding layer, resulting in a concatenated feature map.

[0083] Step 4.3: Input the stitched feature map obtained in step 4.2 into the Inception residual block of the decoding network, and output the feature map with 1 channel through branch 0, which is the absolute phase corresponding to the deformed stripe image of the object under test.

[0084] Finally, it should be noted that the features mentioned and / or shown in the above description of exemplary embodiments of the present invention can be combined in the same or similar manner with one or more other embodiments, combined with features in other embodiments, or substituted for corresponding features in other embodiments. These combined or substituted technical solutions should also be considered to be included within the scope of protection of the present invention.

Claims

1. A deep learning-based end-to-end phase analysis method for single-frame digital stripes, characterized in that, Includes the following steps: Step 1: Build a fringe projection system based on digital twin, acquire deformed fringe images using the phase shift method, calculate the absolute phase of the deformed fringe images using the dual-frequency phase deconvolution algorithm, and construct a deformed fringe-absolute phase dataset. Step 2: Construct the InncepC-Unet network model, which includes an encoder network, a decoder network, and skip connections. The encoder network consists of five encoder layers. The first four encoder layers each consist of an Inception residual block, a coordinate attention block, and a 2×2 max pooling block. The fifth encoder layer consists of an Inception residual block. The decoder network consists of four decoder layers. Each decoder layer consists of a transposed convolution, a concatenation block, and an Inception residual block. Skip connections are used to fuse features between the encoder and decoder layers. The Inception residual block consists of a parallel layer structure composed of branches 0, 1, 2, 3, and 4. Branch 0 consists of convolutional blocks padded with 0 and a batch normalization layer. Branches 1 to 4 each consist of convolutional blocks padded with 1 and a batch normalization layer. The four C / 4 channel outputs of branches 1 to 4 are connected into a single C channel output with a specified dimension of 1, and the C channel output is added to the output of branch 0. The Inception residual block in the encoding network enhances the multi-scale feature extraction capability of the InncepC-Unet network model through multi-branch paths; the coordinate attention block is used to extract the spatial dependence of striped images at different locations in the local feature map, and the coordinate attention block uses one-dimensional pooling in the X and Y directions respectively to replace global two-dimensional pooling, enabling the InncepC-Unet network model to learn the spatial distribution features of the image more accurately and improve the InncepC-Unet network model's attention to the effective target region of the image; 2×2 max pooling can halve the width and height of the extracted feature map, reduce the amount of computation, and improve the training speed of the InncepC-Unet network model; Step 3: Divide the deformed stripe-absolute phase dataset from Step 1 into a training set and a test set according to a certain ratio. Train the InncepC-Unet network model based on the training set and test the trained InncepC-Unet network model based on the test set. Optimize and adjust the relevant parameter settings of the InncepC-Unet network model based on the training and test results. Step 4: Use the InncepC-Unet network model designed in Step 3 to perform absolute phase prediction on the single-frame deformed stripe image of the object under test.

2. The single-frame digital stripe end-to-end phase analysis method according to claim 1, characterized in that, In step 1, the stripe projection system includes a camera, a projector, and the object to be measured. The camera and projector are placed horizontally, and the object to be measured is perpendicular to the projector.

3. The single-frame digital stripe end-to-end phase analysis method according to claim 2, characterized in that, Step 1, the steps for constructing the deformed fringe-absolute phase dataset include: Step 1.1: Render the object to be tested using high-frequency and low-frequency sinusoidal fringes respectively, and then use a camera to capture and save deformed fringe images. The first frame of the high-frequency deformed fringe image is a single-frame deformed fringe image. Step 1.2: For the high-frequency deformed stripe images saved in Step 1.1, excluding the single-frame deformed stripe images, calculate their absolute phase using the dual-frequency phase analysis algorithm and save them; Step 1.3: Replace the object to be tested and repeat the operations of steps 1.1 to 1.

2. Several single-frame deformed stripe images and absolute phases are finally combined to form a deformed stripe-absolute phase dataset.

4. The single-frame digital stripe end-to-end phase analysis method according to claim 3, characterized in that, Step 3, the training steps of the InncepC-Unet network model include: Step 3.1: Normalize the deformed stripe-absolute phase dataset generated in Step 1, and scale the input and output of the InncepC-Unet network model. The input of the InncepC-Unet network model is a single frame deformed stripe image. Step 3.2: Set the relevant parameters for training the InncepC-Unet network model; Step 3.3: Input the predicted phase and the true phase into the loss function of the InncepC-Unet network model. Based on the result of the loss function, train the InncepC-Unet network model until the result of the loss function converges.

5. The single-frame digital stripe end-to-end phase analysis method according to claim 3, characterized in that, In step 3, the deformable stripe-absolute phase dataset is randomly divided into a training set and a test set in a 9:1 ratio, wherein the training set is not visible in the test set.

6. The single-frame digital stripe end-to-end phase analysis method according to claim 1, characterized in that, Step 4, which involves using the InncepC-Unet network model to perform absolute phase prediction on a single frame of deformed stripe images, includes the following steps: Step 4.1: Input the single-frame deformed stripe image of the object under test into the InncepC-Unet network model; Step 4.2: The input single-frame deformed stripe image is first processed by the encoding layer of the encoding network to extract local features and obtain a feature map. Then, the decoding layer of the decoding network performs layer-by-layer transposed convolution to obtain an intermediate feature map. The intermediate feature map is then concatenated with the feature map of the same size in the encoding network to obtain a concatenated feature map. Step 4.3: Input the stitched feature map obtained in step 4.2 into the Inception residual block of the decoding network. The feature map with 1 channel is obtained through branch 0, which is the absolute phase corresponding to the deformed stripe image of the object under test.

Images