Blanket embossing positioning method based on deep learning

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a blanket embossing positioning method based on deep learning. A continuous image sequence of an embossing area and production parameters such as tension and transmission speed are synchronously collected to construct a data set; scanning the image sequence to extract the contour key points of the embossed pattern as positioning anchor points, and delimiting an ROI (Region of Interest) to obtain a coarse positioning result; the production parameters are input into the improved time sequence convolutional network to predict the translation and rotation deformation offset of the blanket, and a coarse positioning result is corrected to generate a positioning mask; and pre-storing geometric parameters of the embossing template, carrying out matching calculation on the positioning mask and the template, and judging whether the standard is reached or not according to a threshold value for positioning. According to the method, multi-source data and time sequence modeling are fused, embossing positioning is achieved through a layered positioning strategy and in combination with intelligent feature processing, the problems that a traditional method is insufficient in positioning precision, and pattern deformation and dislocation are prone to occurring are solved, and the positioning precision and the production efficiency are improved.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and in particular relates to a blanket embossing positioning method based on deep learning. Background Art

[0002] In the blanket production process, embossing is a key step in imparting aesthetics and functionality to product patterns. Its positioning accuracy directly impacts product quality and production efficiency. Traditional blanket embossing positioning methods rely primarily on manual adjustment or rule-based image processing technology, often resulting in insufficient positioning accuracy, pattern deformation, embossing position shift, and misalignment. Summary of the Invention

[0003] In response to the technical problems existing in the above-mentioned background technology, the present invention proposes a blanket embossing positioning method based on deep learning.

[0004] In order to achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0005] S1. Data acquisition: Capture a continuous image sequence of the embossed area at a frame rate synchronized with the felt transmission speed; acquire the production parameters of felt tension, transmission speed, and roller displacement in real time to construct a data set;

[0006] S2, coarse positioning guidance: Rapidly scan the continuous image sequence and extract the key points of the embossed pattern contour as positioning anchor points; delineate the local ROI area based on the positioning anchor points and obtain coarse positioning results for the image within the ROI;

[0007] S3. Production parameter-driven positioning: The felt tension, transmission speed, and roller displacement parameters are input into the improved temporal convolutional network to predict the current felt deformation offset, which includes translation offset and rotation angle offset. Based on the predicted deformation offset, the coarse positioning result is spatially transformed and corrected to generate a positioning mask.

[0008] S4, precise positioning optimization: pre-store the geometric parameters of the embossing template, perform matching calculation on the positioning mask and the template, and judge whether the matching value of the positioning mask meets the standard according to the set matching threshold, which can be used for embossing positioning;

[0009] The specific implementation steps of step S4 are:

[0010] First, a graph structure model of the embossing template and positioning mask is established, where nodes are feature points and edges are geometric constraints between feature points.

[0011] Calculate the matching degree of the two graphs. The matching degree is calculated by first calculating the matching degree of the node position. Where N represents the total number of nodes. Represents the coordinates of the i-th node in the current positioning mask graph, Represents the coordinates of the i-th node in the current template graph; edge matching degree Among them, ε represents the total number of edges, Represents the normalized distance of edge (i, j) in the current positioning mask graph, represents the normalized distance of edge (i, j) in the embossing template graph, Represents the normalized angle of edge (i, j) in the current positioning mask graph, Represents the normalized angle of edge (i, j) in the embossing template graph; the total matching value is obtained by weighted calculation of the matching degree of the node position and the edge matching degree;

[0012] The matching degree is determined according to the set threshold. If so, the current positioning mask is used for embossing positioning. If not, a new positioning mask is generated for matching calculation. The operation is repeated until the matching degree meets the set threshold.

[0013] Preferably, the method for extracting the positioning anchor points in step S2 specifically includes image preprocessing, key point detection, and anchor point screening;

[0014] Image preprocessing: grayscale and Gaussian filtering are performed on continuous image sequences to suppress noise and enhance pattern edge contrast;

[0015] Key point detection: Perform differential operations on continuous image sequences to identify corner points and endpoints of the embossed edge and filter out a set of coarse positioning anchor points;

[0016] Anchor point screening: Based on the spatial distribution density of coarse positioning anchor points, the DBSCAN clustering algorithm is used to remove outliers and retain the cluster center as the positioning anchor point.

[0017] Preferably, the specific implementation of the key point detection includes feature enhancement, corner point detection, endpoint detection and non-redundant screening;

[0018] Feature enhancement: Perform inter-frame difference between the t-th frame image and the t-1-th frame image in a continuous image sequence to calculate the difference image; at the same time, calculate the gradient amplitude of the single frame image, and generate a feature enhancement map by weighted fusion of the difference image and the gradient amplitude;

[0019] Corner detection: Calculate the corner response of the feature enhancement map to generate a corner response matrix. The calculation method is: R 角 (x,y)=det(M)-k(tr(M)) 2 , where R 角 (x, y) represents the corner response matrix, M represents the gradient structure matrix, and k is the empirical coefficient; the pixels in the corner response matrix that are greater than the set threshold are extracted as candidate corner points;

[0020] Endpoint detection: Perform edge detection on the feature enhancement image to generate a binary edge map. Then, extract all edge segments and calibrate the coordinates of the two endpoints of each segment. Define endpoints as pixels with a degree of 1 in the edge segment. Verify the connectivity of the detected endpoints: if the endpoint is connected to at least three consecutive edge pixels in the edge map, it is considered a valid endpoint; otherwise, it is considered as noise removal.

[0021] Non-redundant screening: Merge the candidate corner point set with the valid endpoint set to generate a preliminary anchor point set, and perform non-maximum suppression on the preliminary anchor point set to obtain a non-redundant coarse positioning anchor point set.

[0022] Preferably, the specific implementation of obtaining the coarse positioning result in the image within the ROI in step S2 is as follows:

[0023] First, a convolutional network is used to align the current frame ROI with the corresponding regions of the previous and next frames;

[0024] Using the aligned images, the coordinates of the anchor points are calculated using the centroid algorithm for the positioning anchor points.

[0025] The least square method is used to fit the geometric contour of the embossed pattern edge based on the coordinates of the anchor points to obtain the rough positioning result.

[0026] Preferably, the step S3 of using the improved temporal convolutional network to predict the deformation offset of the current carpet specifically includes:

[0027] The collected felt tension, transmission speed, and roller displacement are spliced ​​into a feature matrix;

[0028] Normalize the feature matrix;

[0029] Multi-layer dilated causal convolution is used, with a convolution kernel size of 3 in each layer, for feature extraction. The feature extraction calculation method is: in, represents the hidden state of the lth layer at time step sj, W l Represents the weight matrix of the lth layer, *d l is the dilated convolution operation, d l represents the dilation factor of layer l, Represented as the l-1 layer at time step sj-d l The hidden state at the time of , is used as the input of the current convolution operation, b l Represents the bias term of the lth layer;

[0030] The residual connection is performed on the output of each layer of dilated causal convolution. The calculation method is: when the number of layers is greater than or equal to 2, in It is the feature output of two layers ago;

[0031] The features of different layers are weighted and aggregated through the attention mechanism to obtain the aggregated features.

[0032] The aggregated features are input into independent prediction heads, which output translation offset (Δx, Δy) and rotation angle offset θ respectively, where Among them, FC θ , FC x , FC y Represents an independent fully connected layer.

[0033] Preferably, in the offset prediction, the loss function of the translation offset adopts the L1 loss function, and the rotation angle adopts the cosine similarity loss function. The expression of the loss function is: L sun =0.7(||Δx pred -Δx gt ||1+||Δy pred -Δy gt ||1)+0.3(1-cos(θ pred -θ gt )), where θ pred , Δx pred , Δy pred represents the predicted value, θ gt , Δx gt , Δy gt Represents the target value.

[0034] Preferably, the specific process of performing spatial transformation correction on the coarse positioning result based on the predicted deformation offset to generate the positioning mask is as follows:

[0035] First, the translation transformation matrix and rotation transformation matrix are constructed using the predicted deformation offset;

[0036] Combining the translation and rotation transformation matrices to form a composite transformation matrix;

[0037] The coordinates of the coarse positioning result are transformed by the composite transformation matrix to obtain the corrected coordinates;

[0038] Based on the corrected coordinates, a closed contour area is drawn in the image, and the pixel points in the closed contour area are assigned a value of 1, and the pixel points outside the area are assigned a value of 0 to generate a positioning mask.

[0039] Compared with the prior art, the advantages and positive effects of the present invention are:

[0040] 1. The present invention constructs a multidimensional data set by synchronously collecting continuous image sequences and production parameters, and uses an improved temporal convolutional network to extract temporal features of the production parameters, thereby achieving accurate prediction of the translational and rotational deformation of the blanket, solving the positioning deviation problem caused by ignoring dynamic deformation in traditional methods.

[0041] 2. Through a layered processing flow, image key point detection and ROI area analysis are used to quickly obtain coarse positioning results. Spatial transformation correction is then performed based on the deformation offset predicted by production parameters. Finally, geometric constraints are verified using a graph structure matching model and template. This significantly improves positioning accuracy and avoids the pattern deformation and misalignment problems caused by a single processing flow in traditional rule-based methods.

[0042] 3. In the key point detection stage, the feature enhancement technology of inter-frame difference and gradient amplitude fusion is combined with corner point response calculation and endpoint connectivity verification to effectively suppress noise interference such as hair jitter and light fluctuation in blanket transmission; at the same time, the DBSCAN clustering algorithm is used to screen positioning anchor points, reduce the influence of outliers, and improve the reliability of feature point distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0044] Figure 1 This is a flowchart of the overall structure of a blanket embossing positioning method based on deep learning; DETAILED DESCRIPTION

[0045] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0046] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways than those described herein. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0047] In the embossing process of large-scale blanket production enterprises, the existing production line adopts a positioning method that combines manual visual inspection with rule-based image processing, which often leads to insufficient positioning accuracy, pattern deformation, embossing position offset and misalignment. To address these problems, the present invention proposes a blanket embossing positioning method based on deep learning, the specific steps are as follows: Figure 1 shown.

[0048] First, traditional methods only consider single-frame images and lack temporal modeling of the dynamic deformation process of the felt. This method first acquires data, capturing a continuous image sequence of the embossed area at a frame rate synchronized with the felt's transmission speed. Simultaneously, production parameters such as felt tension, transmission speed, and roller displacement are acquired in real time to construct a dataset.

[0049] Then, coarse positioning guidance is performed: the continuous image sequence is quickly scanned, and the key points of the embossed pattern contour are extracted as positioning anchor points; the local ROI area is delineated based on the positioning anchor points, and the coarse positioning result is obtained within the image within the ROI.

[0050] Specifically, the method for extracting the positioning anchor points specifically includes image preprocessing, key point detection, and anchor point screening. Image preprocessing is to grayscale and Gaussian filter the continuous image sequence to suppress noise and enhance the pattern edge contrast. Specifically, the continuous image sequence is first grayscaled to convert the color image into a grayscale image, weakening the color information to retain the light and dark structure and reduce the computational complexity. Then Gaussian filtering is performed, and the image is convolved with a Gaussian kernel of a specific size. By weighted averaging the pixel neighborhood, the noise generated by the fluff jitter, light fluctuations, etc. in the blanket transmission is suppressed, while retaining the low-frequency features of the pattern edge to obtain the processed image. Key point detection is to perform differential operations on the continuous image sequence, identify the corner points and endpoints of the embossed edge, and screen out a set of coarse positioning anchor points. Specifically, the specific implementation of key point detection includes feature enhancement, corner point detection, endpoint detection, and non-redundant screening. Feature enhancement: perform inter-frame difference between the t-th frame image and the t-1-th frame image in the continuous image sequence to calculate the difference image; at the same time, calculate the gradient amplitude of the single frame image, and generate the feature enhancement image by weighted fusion of the difference image and the gradient amplitude; corner detection: calculate the corner response of the feature enhancement image to generate the corner response matrix, which is calculated as follows: R 角 (x,y)=det(M)-k(tr(M)) 2 , where R 角(x,y) represents the corner response matrix, M represents the gradient structure matrix, and k is the empirical coefficient. Pixels in the corner response matrix that are greater than a set threshold are extracted as candidate corner points. Endpoint detection is performed on the feature enhancement image to generate a binary edge map. All edge segments are then extracted and the coordinates of the two endpoints of each segment are calibrated. Endpoints are defined as pixels with a degree of 1 in the edge segment. Connectivity verification is performed on the detected endpoints: if an endpoint is connected to at least three consecutive edge pixels in the edge map, it is considered a valid endpoint; otherwise, it is treated as noise and removed. Finally, non-redundant filtering is performed, merging the candidate corner point set with the valid endpoint set to generate a preliminary anchor point set. Non-maximum suppression is performed on the preliminary anchor point set to obtain a non-redundant coarse anchor point set. After selecting the group anchor points, the anchor points are screened. Based on the spatial distribution density of the coarse anchor points, the DBSCAN clustering algorithm is used to remove outliers, retaining the cluster center as the anchor point. Specifically, based on the coordinates of the coarse anchor points, the DBSCAN clustering algorithm was used to select anchor points. The algorithm set a neighborhood radius of eps = 15 pixels and a minimum number of samples, min_samples = 5. The anchor points were then traversed and clustered. Areas with ≥ 5 anchor points within the eps range were considered valid clusters, while the rest were considered outliers. After removing the outliers, the centers of each cluster were used as the anchor points for positioning.

[0051] Then, based on the positioning anchor point, the local ROI area is delineated, and the coarse positioning result is obtained for the image within the ROI. This process is to first use the convolutional network to perform feature matching and geometric transformation on the ROI area of ​​the current frame and the corresponding area of ​​the previous and next adjacent frames, compensate for the displacement difference caused by the blanket transmission, and achieve sub-pixel alignment. Then, for the aligned image, the centroid algorithm is used to calculate the precise coordinates of the positioning anchor point, and its geometric center is determined by weighted averaging the pixel values ​​of the anchor point area. Finally, based on the anchor point coordinates, the least squares method is used to perform geometric fitting on the edge of the embossed pattern. For example, the outline of the maple leaf is approximated as a polynomial curve or a piecewise straight line, and the fitting error is less than the pixel of the set threshold, thereby generating a coarse positioning result containing the coordinates of the pattern outline, providing an initial benchmark for subsequent deformation correction.

[0052] Next, the blanket tension, transmission speed, and roller displacement parameters are input into the improved temporal convolutional network to predict the current blanket deformation offset, which includes translation offset and rotation angle offset. The improved temporal convolutional network is specifically implemented to predict the deformation offset by concatenating the collected blanket tension, transmission speed, and roller displacement into a feature matrix; normalizing the feature matrix; and using multi-layer dilated causal convolution with a convolution kernel size of 3 per layer for feature extraction. The feature extraction calculation method is as follows: in, represents the hidden state of the lth layer at time step sj, W l represents the weight matrix of layer l, is the dilated convolution operation, d l represents the dilation factor of layer l, Represented as the l-1 layer at time step sj-d l The hidden state at the time of , is used as the input of the current convolution operation, b l Represents the bias term of the lth layer; the output of each layer of dilated causal convolution is residually connected, and the calculation method is: when the number of layers is greater than or equal to 2, in is the feature output of the previous two layers; then the features of different layers are weighted and aggregated through the attention mechanism to obtain the aggregated features Finally, the aggregated features are input into independent prediction heads, which output the translation offset (Δx, Δy) and rotation angle offset θ respectively, where Among them, FC θ , FC x , FC y Represents an independent fully connected layer. Among them, the loss function of the translation offset adopts the L1 loss function, and the rotation angle adopts the cosine similarity loss function. The expression of the loss function is: L sun =0.7(||Δx pred -Δx gt ||1+||Δy pred -Δy gt ||1)+0.3(1-cos(θ pred -θ gt )), where θ pred , Δx pred , Δy pred represents the predicted value, θ gt , Δx gt , Δy gt Represents the target value.

[0053] The coarse positioning results are spatially transformed and corrected based on the predicted deformation offset to generate a positioning mask. First, a translation transformation matrix and a rotation transformation matrix are constructed using the predicted deformation offset. The translation and rotation transformation matrices are combined to form a composite transformation matrix. The coordinates of the coarse positioning results are transformed using the composite transformation matrix to obtain corrected coordinates. Based on the corrected coordinates, a closed contour region is drawn in the image, and pixels within the closed contour region are assigned a value of 1, while pixels outside the region are assigned a value of 0 to generate a positioning mask. Specifically, a translation transformation matrix and a rotation transformation matrix are constructed based on the blanket deformation parameters predicted by the improved temporal convolutional network. The translation transformation matrix is ​​used to shift the coordinates of the coarse positioning results by a specified number of pixels horizontally and vertically on the image plane to compensate for positional offsets during blanket transport. The rotation transformation matrix is ​​used to rotate the coordinates around the image origin by a specified angle to correct for rotational deformation caused by factors such as uneven blanket tension. Subsequently, the translation and rotation matrices are combined into a composite transformation matrix, and matrix multiplication is used to achieve superposition correction of the two deformations. The order of the composite transformation is to perform a rotation transformation first, followed by a translation transformation. This ensures that the rotation operation is based on the original coordinate origin, avoiding correction errors caused by translating first and then rotating. Next, the embossed pattern contour coordinates obtained from the rough positioning are input into the composite transformation matrix. Each coordinate point is spatially transformed through matrix operations to obtain the corrected coordinate values. This process can be understood as: contour points that originally deviated from the ideal position due to blanket deformation are reversed according to the predicted offset to return them to the theoretically correct position, thereby compensating for the blanket's translational and rotational deformation. Based on the corrected contour coordinates, a closed contour area is drawn in the image. The specific operation includes: using an interpolation algorithm to connect adjacent contour points to form a continuous closed curve, ensuring that the contour has no breaks or gaps. Then, the area inside the closed contour is pixel-labeled: all pixels within the contour are assigned a value of 1 (representing the valid positioning area), and pixels outside the contour are assigned a value of 0 (representing the background area), generating a binary positioning mask.

[0054] Finally, perform precise positioning optimization, use the pre-stored geometric parameters of the embossing template, match the positioning mask with the template, and determine whether the matching value of the positioning mask meets the standard based on the set matching threshold, which can be used for embossing positioning. The specific implementation steps are: first, establish a graph structure model of the embossing template and the positioning mask, where the nodes are feature points and the edges are the geometric constraint relationships between feature points. The operation of extracting nodes and edges from the positioning mask is to extract closed contours from the binary image of the positioning mask to form a feature point set containing coordinates and types as nodes. Secondly, the geometric constraint relationship of the edge is constructed by connecting adjacent nodes according to k nearest neighbors, calculating the normalized distance and normalized angle of each edge, and forming an edge set containing adjacency relationships and geometric parameters. Then calculate the matching degree of the two graphs. The matching degree calculation is to first calculate the matching degree of the node position. Where N represents the total number of nodes. Represents the coordinates of the i-th node in the current positioning mask graph, Represents the coordinates of the i-th node in the current template graph; edge matching degree Among them, ε represents the total number of edges, Represents the normalized distance of edge (i, j) in the current positioning mask graph, represents the normalized distance of edge (i, j) in the embossing template graph, Represents the normalized angle of edge (i, j) in the current positioning mask graph, Represents the normalized angle of edge (i, j) in the embossing template graph; performs weighted calculation on the node position matching degree and edge matching degree to obtain the total matching degree value; determines whether the matching degree value meets the standard according to the set threshold; if so, uses the current positioning mask for embossing positioning; otherwise, regenerates the positioning mask for matching calculation, and repeats the operation until the matching degree meets the set threshold.

[0055] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any person skilled in the art may utilize the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes for application in other fields. However, any simple modification, equivalent change, and modification of the above embodiments made in accordance with the technical essence of the present invention without departing from the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A blanket embossing positioning method based on deep learning, characterized in that: The following steps are involved: S1. Data acquisition: Capture a continuous image sequence of the embossed area at a frame rate synchronized with the felt transmission speed; acquire the production parameters of felt tension, transmission speed, and roller displacement in real time to construct a data set; S2, coarse positioning guidance: Rapidly scan the continuous image sequence and extract the key points of the embossed pattern contour as positioning anchor points; Delineate the local ROI area based on the positioning anchor point, and obtain the rough positioning result within the image within the ROI; S3. Production parameter-driven positioning: The felt tension, transmission speed, and roller displacement parameters are input into the improved temporal convolutional network to predict the current felt deformation offset, which includes translation offset and rotation angle offset. Based on the predicted deformation offset, the coarse positioning result is spatially transformed and corrected to generate a positioning mask. S4, precise positioning optimization: pre-store the geometric parameters of the embossing template, perform matching calculation on the positioning mask and the template, and judge whether the matching value of the positioning mask meets the standard according to the set matching threshold, which can be used for embossing positioning; The specific implementation steps of step S4 are: First, a graph structure model of the embossing template and positioning mask is established, where nodes are feature points and edges are geometric constraints between feature points. Calculate the matching degree of the two graphs. The matching degree is calculated by first calculating the matching degree of the node position. Where N represents the total number of nodes. Represents the coordinates of the i-th node in the current positioning mask graph, Represents the coordinates of the i-th node in the current template graph; edge matching degree Among them, ε represents the total number of edges, Represents the normalized distance of edge (i, j) in the current positioning mask graph, represents the normalized distance of edge (i, j) in the embossing template graph, Represents the normalized angle of edge (i, j) in the current positioning mask graph, Represents the normalized angle of edge (i, j) in the embossing template graph; the total matching value is obtained by weighted calculation of the matching degree of the node position and the edge matching degree; The matching degree is determined according to the set threshold. If so, the current positioning mask is used for embossing positioning. If not, a new positioning mask is generated for matching calculation. The operation is repeated until the matching degree meets the set threshold.

2. A method for positioning felt embossing based on deep learning according to claim 1, characterized in that: The method for extracting the positioning anchor points in step S2 specifically includes image preprocessing, key point detection, and anchor point screening; Image preprocessing: grayscale and Gaussian filtering are performed on continuous image sequences to suppress noise and enhance pattern edge contrast; Key point detection: Perform differential operations on continuous image sequences to identify corner points and endpoints of the embossed edge and filter out a set of coarse positioning anchor points; Anchor point screening: Based on the spatial distribution density of coarse positioning anchor points, the DBSCAN clustering algorithm is used to remove outliers and retain the cluster center as the positioning anchor point.

3. The method for positioning felt embossing based on deep learning according to claim 2, characterized in that: The specific implementation of the key point detection includes feature enhancement, corner point detection, endpoint detection and non-redundant screening; Feature enhancement: Perform inter-frame difference between the t-th frame image and the t-1-th frame image in a continuous image sequence to calculate the difference image; at the same time, calculate the gradient amplitude of the single frame image, and generate a feature enhancement map by weighted fusion of the difference image and the gradient amplitude; Corner detection: Calculate the corner response of the feature enhancement map to generate a corner response matrix. The calculation method is: R 角 (x,y)=det(M)-k(tr(M)) 2 , where R 角 (x, y) represents the corner response matrix, M represents the gradient structure matrix, and k is the empirical coefficient; the pixels in the corner response matrix that are greater than the set threshold are extracted as candidate corner points; Endpoint detection: Perform edge detection on the feature enhancement image to generate a binary edge map. Then, extract all edge segments and calibrate the coordinates of the two endpoints of each segment. Define endpoints as pixels with a degree of 1 in the edge segment. Verify the connectivity of the detected endpoints: if the endpoint is connected to at least three consecutive edge pixels in the edge map, it is considered a valid endpoint; otherwise, it is considered as noise removal. Non-redundant screening: Merge the candidate corner point set with the valid endpoint set to generate a preliminary anchor point set, and perform non-maximum suppression on the preliminary anchor point set to obtain a non-redundant coarse positioning anchor point set.

4. The method for positioning felt embossing based on deep learning according to claim 1, characterized in that: The specific implementation of obtaining the coarse positioning result in the image within the ROI in step S2 is as follows: First, a convolutional network is used to align the current frame ROI with the corresponding regions of the previous and next frames; Using the aligned images, the coordinates of the anchor points are calculated using the centroid algorithm for the positioning anchor points. The least square method is used to fit the geometric contour of the embossed pattern edge based on the coordinates of the anchor points to obtain the rough positioning result.

5. The method for positioning felt embossing based on deep learning according to claim 1, characterized in that: In step S3, the improved temporal convolutional network is used to predict the deformation offset of the current blanket, specifically including: The collected felt tension, transmission speed, and roller displacement are spliced ​​into a feature matrix; Normalize the feature matrix; Multi-layer dilated causal convolution is used, with a convolution kernel size of 3 in each layer, for feature extraction. The feature extraction calculation method is: in, represents the hidden state of the lth layer at time step sj, W l represents the weight matrix of layer l, is the dilated convolution operation, d l represents the dilation factor of layer l, Represented as the l-1 layer at time step sj-d l The hidden state at the time of , is used as the input of the current convolution operation, b l Represents the bias term of the lth layer; The residual connection is performed on the output of each layer of dilated causal convolution. The calculation method is: when the number of layers is greater than or equal to 2, in It is the feature output of two layers ago; The features of different layers are weighted and aggregated through the attention mechanism to obtain the aggregated features. The aggregated features are input into independent prediction heads, which output translation offset (Δx, Δy) and rotation angle offset θ respectively, where Among them, FC θ , FC x , FC y Represents an independent fully connected layer.

6. A method for positioning felt embossing based on deep learning according to claim 5, characterized in that: In the offset prediction, the loss function of the translation offset adopts the L1 loss function, and the rotation angle adopts the cosine similarity loss function. The expression of the loss function is: L sun =0.7(||Δx pred -Δx gt ||1+||Δy pred -Δy gt ||1)+0.3(1-cos(θ pred -θ gt )), where θ pred , Δx pred , Δy pred represents the predicted value, θ gt , Δx gt , Δy gt Represents the target value.

7. The method for positioning felt embossing based on deep learning according to claim 5, characterized in that: The specific process of performing spatial transformation correction on the coarse positioning result based on the predicted deformation offset and generating the positioning mask is as follows: First, the translation transformation matrix and rotation transformation matrix are constructed using the predicted deformation offset; Combining the translation and rotation transformation matrices to form a composite transformation matrix; The coordinates of the coarse positioning result are transformed by the composite transformation matrix to obtain the corrected coordinates; Based on the corrected coordinates, a closed contour area is drawn in the image, and the pixel points in the closed contour area are assigned a value of 1, and the pixel points outside the area are assigned a value of 0 to generate a positioning mask.

Images