A deep learning-based blanket embossing positioning method

Abstract

The present application belongs to the technical field of image processing, and particularly relates to a blanket embossing positioning method based on deep learning. A data set is constructed by synchronously collecting a continuous image sequence of an embossing area and production parameters such as tension and transmission speed; key points of an embossing pattern contour are extracted from the image sequence as positioning anchor points, and a ROI area is demarcated to obtain a coarse positioning result; production parameters are input into an improved time sequence convolution network to predict the translation and rotation deformation offset of the blanket, and the coarse positioning result is corrected to generate a positioning mask; the geometric parameters of a pre-stored embossing template are matched with the positioning mask for calculation, and whether the positioning is up to the standard is judged according to a threshold value. The present application combines multi-source data and time sequence modeling, realizes embossing positioning through a hierarchical positioning strategy and intelligent feature processing, and solves the problems of insufficient positioning accuracy, pattern deformation and mispositioning of traditional methods, thereby improving positioning accuracy and production efficiency.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and particularly relates to a blanket embossing positioning method based on deep learning. BACKGROUND

[0002] In the blanket production process, the embossing process is a key link to give the product pattern beauty and functionality, and its positioning accuracy directly affects the product quality and production efficiency. The traditional blanket embossing positioning method mainly relies on manual adjustment or rule-based image processing technology, which often has insufficient positioning accuracy, and has problems such as pattern deformation, embossing position deviation and misplacement. SUMMARY

[0003] The present application proposes a blanket embossing positioning method based on deep learning to solve the technical problems in the above background technology.

[0004] In order to achieve the above purpose, the technical scheme adopted by the present application comprises the following steps:

[0005] S1, data acquisition: acquiring a continuous image sequence of the embossing area at a frame rate synchronized with the blanket transmission speed; simultaneously acquiring real-time blanket tension, transmission speed, roller shaft displacement production parameters, and constructing a data set;

[0006] S2, coarse positioning guide: rapidly scanning the continuous image sequence, extracting the embossing pattern contour key points as positioning anchor points; based on the positioning anchor points, defining a local ROI area, and obtaining a coarse positioning result in the image in the ROI;

[0007] S3, production parameter driven positioning: inputting the blanket tension, transmission speed and roller shaft displacement parameters into an improved time sequence convolution network to predict the deformation offset of the current blanket, the deformation offset including a translation offset and a rotation angle offset; based on the predicted deformation offset, performing spatial transformation correction on the coarse positioning result to generate a positioning mask;

[0008] S4, fine positioning optimization: pre-storing the geometric parameters of the embossing template, matching and calculating the positioning mask with the template, and judging whether the positioning mask meets the matching threshold 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 the positioning mask is established, in which the nodes are feature points and the edges are geometric constraint relationships between the feature points;

[0011] The matching degree of the two graphs is calculated, which is calculated as the matching degree of the node positions where N represents the total number of nodes, and represents the matching degree of the current positioning mask graph The coordinates of each node, Represents the current template image number The coordinates of each node; edge matching degree ,in, Represents the total number of edges. Represents the edge in the current location mask image Normalized distance, Represents the edge of the embossing template diagram Normalized distance, Represents the edge in the current location mask image From the perspective of normalization, Represents the edge of the embossing template diagram The normalized angle; the matching degree of the node position and the edge matching degree are weighted and calculated to obtain the total matching degree value;

[0012] The system determines whether the matching degree meets the set threshold. If it does, the current positioning mask is used for embossing positioning. Otherwise, the positioning mask is regenerated and the matching calculation is performed. This process is repeated until the matching degree meets the set threshold.

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

[0014] Image preprocessing: The continuous image sequence is converted to grayscale and Gaussian filtered to suppress noise and enhance the contrast of pattern edges;

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

[0016] Anchor point selection: Based on the spatial distribution density of coarse positioning anchor points, outliers are removed using the DBSCAN clustering algorithm, and cluster centers are retained as positioning anchor points.

[0017] Preferably, the key point detection is specifically implemented by including feature enhancement, corner detection, endpoint detection, and non-redundant filtering.

[0018] Feature enhancement: Perform inter-frame difference analysis between the t-th frame and the (t-1)-th frame in a continuous image sequence to calculate the difference image; simultaneously calculate the gradient magnitude for a single frame image, and generate a feature enhancement map by weighted fusion of the difference image and the gradient magnitude.

[0019] Corner detection: Corner response calculation is performed on the feature enhancement map to generate a corner response matrix. The calculation method is as follows: ,in The corner response matrix is ​​represented by M, the gradient structure matrix is ​​represented by k, and the empirical coefficient is k. Pixels in the corner response matrix that are greater than a set threshold are extracted as candidate corners.

[0020] Endpoint detection: Edge detection is performed on the feature enhancement map to generate a binary edge map. Then, all edge segments are extracted, and the coordinates of the two endpoints of each segment are calibrated. The endpoint is defined as a pixel with a degree of 1 in the edge segment. The connectivity of the detected endpoints is verified: if the endpoint is connected to at least 3 consecutive edge pixels in the edge map, it is determined to be a valid endpoint; otherwise, it is considered noise and is removed.

[0021] Non-redundant screening: The candidate corner point set is merged with the valid endpoint set to generate a preliminary anchor point set. Non-maximum suppression processing is applied to the preliminary anchor point set to obtain a coarse positioning anchor point set without redundancy.

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

[0023] First, a convolutional network is used to align the ROI of the current frame with the corresponding regions of the adjacent frames before and after it;

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

[0025] Based on the coordinates of the anchor points, the least squares method is used to fit the geometric contour of the embossed pattern edge to obtain a coarse positioning result.

[0026] Preferably, step S3, which uses an improved temporal convolutional network to predict the current deformation offset of the blanket, specifically includes:

[0027] The collected data on blanket tension, transmission speed, and roller displacement are spliced ​​together to form a feature matrix.

[0028] Standardize the feature matrix;

[0029] Multi-layer dilated causal convolution, with each layer having a kernel size of 3, is used for feature extraction. The feature extraction calculation method is as follows: ,in, Indicates the first The hidden state of the layer at time step sj. Representing the The weight matrix of the layer, For dilated convolution operations, Representing the The expansion factor of the layer Represented as the first Layer in time step The hidden state at that time is used as the input for the current convolution operation. Representing the Layer bias terms;

[0030] Residual connection is performed on the output of each layer of dilated causal convolution, and the calculation method is as follows: when the number of layers is greater than or equal to 2, , wherein is the feature output of two layers before;

[0031] The aggregated features are obtained by weighting and aggregating the features of different layers through an attention mechanism .

[0032] The aggregated features are input into independent prediction heads, and the translation offset and the rotation angle offset are output respectively, wherein , , , wherein, represents an independent fully connected layer.

[0033] As a preferred, in the offset prediction, the loss function of the translation offset adopts an L1 loss function, and the cosine similarity loss function is adopted for the rotation angle, and the expression of the loss function is as follows: , wherein, represents a predicted value, represents a target value.

[0034] As a preferred, the spatial transformation correction is performed on the coarse positioning result based on the predicted deformation offset to generate the positioning mask, and the specific process of the implementation is as follows:

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

[0036] The translation and rotation transformation matrices are combined to form a composite transformation matrix;

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

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

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

[0040] 1. The present application realizes accurate prediction of the translation and rotation deformation of the blanket by synchronously collecting a continuous image sequence and production parameters, constructing a multi-dimensional data set, and using an improved time sequence convolution network to extract time sequence features of the production parameters.

[0041] 2、Through the hierarchical processing flow, the coarse positioning result is quickly obtained through image key point detection and ROI region analysis, then the spatial transformation correction is carried out combined with the deformation offset predicted by the production parameter, finally the geometric constraint verification is carried out through the graph structure matching model and the template, which significantly improves the positioning accuracy and avoids the pattern deformation and misplacement problems caused by single processing flow in the traditional rule-based method.

[0042] 3、In the key point detection link, through the feature enhancement technology of inter-frame difference and gradient amplitude fusion, combined with the corner response calculation and end point connectivity verification, the noise interference such as fluff shaking and light fluctuation in the blanket transmission is effectively suppressed; at the same time, the DBSCAN clustering algorithm is used to filter the positioning anchor points, which reduces the influence of outliers and improves the reliability of the feature point distribution. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0044] Figure 1 It is a whole structure flow chart of a blanket embossing positioning method based on deep learning; DETAILED DESCRIPTION

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0046] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the present application is not limited to the specific embodiments disclosed in the following description.

[0047] In the embossing process link of large-scale blanket production enterprises, the existing production line adopts the positioning method combined with manual visual detection and rule-based image processing, which often causes insufficient positioning accuracy, pattern deformation, embossing position offset and misplacement. In view of these problems, the present application proposes a blanket embossing positioning method based on deep learning, and the specific steps are as shown in Figure 1 .

[0048] Firstly, the traditional method only considers single frame image, lacking of time sequence modeling of the dynamic deformation process of the blanket. The present application firstly collects data, collects continuous image sequence of the embossed area at a frame rate synchronized with the transmission speed of the blanket; simultaneously, real-time obtains the production parameters of the blanket tension, transmission speed, roller shaft displacement, and constructs the data set.

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

[0050] Specifically, the extraction method of the positioning anchor point specifically includes image preprocessing, key point detection, and anchor point screening. The image preprocessing is to perform gray scale and Gaussian filtering processing on the continuous image sequence, suppress noise and enhance the contrast of the pattern edge. Specifically, the continuous image sequence is first processed by gray scale to convert the color image into a gray scale image, weaken the color information to retain the light and dark structure, and reduce the computational complexity. Then, Gaussian filtering is performed, the image is convolved with a specific size of Gaussian kernel, the noise generated by the fluff shaking, light fluctuation and the like in the transmission of the blanket is suppressed by weighted average of the pixel neighborhood, and the low frequency features of the pattern edge are retained, to obtain the processed image. The key point detection is to perform difference operation on the continuous image sequence, identify the corner points and end points of the embossed edge, and screen out the coarse positioning anchor point set. Specifically, the specific implementation of the key point detection includes feature enhancement, corner point detection, end point detection, and non-redundant screening. The feature enhancement is to perform inter-frame difference between the t-th frame image and the t-1-th frame image in the continuous image sequence, calculate the difference image; at the same time, the gradient amplitude of the single frame image is calculated, and the feature enhancement image is generated by weighted fusion of the difference image and the gradient amplitude; the corner point detection is to calculate the corner point response of the feature enhancement image, generate the corner point response matrix, and the calculation method is: wherein The corner response matrix is ​​represented by M, the gradient structure matrix, and k is an empirical coefficient. Pixels with a value greater than a set threshold in the corner response matrix are extracted as candidate corners. Endpoint detection is performed on the feature enhancement map to generate a binary edge map. Then, all edge segments are extracted, and the coordinates of the two endpoints of each segment are calibrated. An endpoint is defined as a pixel 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 3 consecutive edge pixels in the edge map, it is considered a valid endpoint; otherwise, it is considered noise and discarded. Finally, non-redundant filtering is performed by merging the candidate corner set and the valid endpoint set to generate a preliminary anchor set. Non-maximum suppression is applied to the preliminary anchor set to obtain a coarse anchor set without redundancy. After selecting a group of anchors, the anchors are filtered. Based on the spatial distribution density of the coarse anchors, outliers are removed using the DBSCAN clustering algorithm, and cluster centers are retained as anchors. Specifically, based on the coarse positioning anchor point coordinates, the DBSCAN clustering algorithm is used to filter anchor points: the neighborhood radius eps = 15 pixels and the minimum number of samples min_samples = 5 are set. The anchor points are traversed and clustered. Regions with ≥5 anchor points within the eps range are set as valid clusters, and the rest are outliers. After removing outliers, the center of each cluster is taken as the positioning anchor point.

[0051] Next, local Regions of Interest (ROIs) are delineated based on the anchor points, and coarse localization results are obtained within the ROI image. This process first uses a convolutional network to perform feature matching and geometric transformation on the ROI region of the current frame and the corresponding regions of the preceding and following frames, compensating for displacement differences caused by blanket transmission and achieving sub-pixel alignment. Then, for the aligned image, the centroid algorithm is used to calculate the precise coordinates of the anchor points, and the geometric center is determined by weighted averaging of the pixel values ​​of the anchor point region. Finally, based on the anchor point coordinates, the least squares method is used to geometrically fit the edges of the embossed pattern, for example, approximating the maple leaf outline as a polynomial curve or piecewise straight line. Pixels with a fitting error less than a set threshold are used to generate coarse localization results containing the pattern outline coordinates, providing an initial benchmark for subsequent deformation correction.

[0052] Next, the blanket tension, transmission speed, and roller displacement parameters are input into an improved temporal convolutional network to predict the current deformation offset of the blanket. This deformation offset includes translational offset and rotational angle offset. Specifically, the improved temporal convolutional network predicts the deformation offset by concatenating the collected blanket tension, transmission speed, and roller displacement into a feature matrix; standardizing the feature matrix; and using multi-layer dilated causal convolutions with a kernel size of 3 for feature extraction. The feature extraction is calculated as follows: ,in, Indicates the first The hidden state of the layer at time step sj. Representing the a weight matrix of the layer, for dilated convolution operation, representing the first dilating factor of the layer, representing the first hidden state of the layer at the time step as the input of the current convolution operation, representing the first bias term of the layer; the output of each dilated causal convolution is connected in residual connection, and the calculation method is as follows: when the number of layers is greater than or equal to 2, wherein is the feature output before two layers; then the features of different layers are weighted and aggregated through an attention mechanism to obtain aggregated features ; finally, the aggregated features are input into independent prediction heads to respectively output translation offset and rotation angle offset , wherein , , wherein, representing 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, and the expression of the loss function is as follows: wherein, representing the predicted value, representing the target value.

[0053] The coarse positioning result is corrected by spatial transformation based on the predicted deformation offset to generate a positioning mask. First, a translation transformation matrix and a rotation transformation matrix are constructed by the predicted deformation offset; the translation and rotation transformation matrices are combined to form a composite transformation matrix; the coordinates of the coarse positioning result are transformed by the composite transformation matrix to obtain corrected coordinates; based on the corrected coordinates, a closed contour region is drawn in the image, the pixel points in the closed contour region are assigned a value of 1, and the pixel points outside the region are assigned a value of 0 to generate a positioning mask. Specifically, first, the translation transformation matrix and the rotation transformation matrix are constructed according to the blanket deformation parameters predicted by the improved temporal convolutional network. The translation transformation matrix is used to move the coordinates of the coarse positioning result in the image plane along the horizontal and vertical directions by a specified number of pixels to compensate for the position offset in the blanket transmission process; the rotation transformation matrix is used to rotate the coordinates around the image origin by a specified angle to correct the rotational deformation of the blanket caused by uneven tension and other factors. Then, the translation matrix and the rotation matrix are combined into a composite transformation matrix to realize the superposition correction of the two deformations by matrix multiplication. The order of the composite transformation is to perform the rotation transformation first and then the translation transformation, which ensures that the rotation operation is based on the original coordinate origin and avoids the correction deviation caused by translation first and then rotation. Next, the embossed pattern contour coordinates obtained by coarse positioning are input into the composite transformation matrix, and the spatial transformation of each coordinate point is realized by matrix operation to obtain the corrected coordinate values. This process can be understood as follows: the contour points that deviate from the ideal position due to blanket deformation are adjusted in the opposite direction according to the predicted offset to return to the correct position in theory, thereby compensating for the translation and rotation deformation of the blanket. Based on the corrected contour coordinates, a closed contour region is drawn in the image. The specific operation includes: connecting adjacent contour points using an interpolation algorithm to form a continuous closed curve, ensuring that the contour has no breakpoints or gaps. Then, the pixels in the closed contour region are marked: all pixel points inside the contour are assigned a value of 1 (representing the effective positioning region), and pixel points outside the contour are assigned a value of 0 (representing the background region) to generate a binary positioning mask.

[0054] Finally, fine positioning optimization is performed, the positioning mask is matched with the pre-stored embossing template to calculate the matching degree, and whether the positioning mask meets the matching threshold value for embossing positioning is determined. The specific implementation steps are as follows: first, a graph structure model of the embossing template and the positioning mask is established, where the nodes are feature points and the edges are geometric constraint relationships between the feature points. For the positioning mask, the operation of extracting nodes and edges is to extract a closed contour from the binary image of the positioning mask to form a feature point set containing coordinates and types as nodes. Then, the geometric constraint relationship of the edge is to connect adjacent nodes according to the k-nearest neighbor, calculate the normalized distance and normalized angle of each edge, and form an edge set containing adjacency relationship and geometric parameters. Then, the matching degree of the two graphs is calculated where N represents the total number of nodes, coordinates of the i-th node of the current positioning mask map, coordinates of the i-th node of the current template map; edge matching degree wherein, total number of edges, normalized distance of the edge in the current positioning mask map, normalized distance of the edge in the embossing template map, normalized angle of the edge in the current positioning mask map, normalized angle of the edge in the embossing template map; The total matching degree value is calculated by weighting the matching degree of the node position. Whether the matching degree value meets the set threshold is determined. If yes, the current positioning mask is used for embossing positioning. If not, the positioning mask is regenerated for matching calculation. The operation is repeated until the matching degree meets the set threshold.

[0055] The above is only a preferred embodiment of the present application, and is not intended to limit the present application in other forms. Any skilled person in the art can modify or change the above disclosed technical content to equivalent embodiments applied to other fields. However, any simple modification, equivalent change and modification of the above embodiments made according to the technical essence of the present application without departing from the technical solution content of the present application still falls within the protection scope of the present application.​​​​

Claims

1. A deep learning-based blanket embossing positioning method, characterized in that, The method comprises the following steps: S1, data acquisition: acquiring a continuous image sequence of the embossed area at a frame rate synchronized with the transmission speed of the blanket; simultaneously, real-time acquisition of the blanket tension, transmission speed, and roller shaft displacement production parameters to construct a data set; S2, coarse positioning guidance: rapid scanning of the continuous image sequence to extract the key points of the embossed pattern contour as positioning anchor points; Based on the positioning anchor points, a local ROI region is demarcated, and a coarse positioning result is obtained in the image in the ROI; S3, production parameter driven positioning: inputting the blanket tension, transmission speed, and roller shaft displacement parameters into the improved time sequence convolution network to predict the deformation offset of the current blanket, the deformation offset including a translation offset and a rotation angle offset; based on the predicted deformation offset, the coarse positioning result is spatially transformed and corrected to generate a positioning mask; S4, fine positioning optimization: pre-storing the geometric parameters of the embossed template, matching and calculating the positioning mask with the template, and judging whether the positioning mask meets the matching threshold according to the set matching threshold to be used for embossing positioning; The specific implementation steps of the step S4 are: First, a graph structure model of the embossed template and the positioning mask is established, wherein the nodes are feature points and the edges are the geometric constraint relationship between the feature points; Calculate the matching degree between the two graphs, wherein the matching degree is calculated by first calculating the matching degree of the node positions. Where N represents the total number of nodes. Represents the current location mask image. The coordinates of each node, Represents the current template image number The coordinates of each node; edge matching degree ,in, Represents the total number of edges. Represents the edge in the current location mask image Normalized distance, Represents the edge of the embossing template diagram Normalized distance, Represents the edge in the current location mask image From the perspective of normalization, Represents the edge of the embossing template diagram The normalized angle; the matching degree of the node position and the edge matching degree are weighted and calculated to obtain the total matching degree value; According to the set threshold, it is judged whether the matching degree value meets the standard, if yes, the current positioning mask is used for embossing positioning, if not, the positioning mask is regenerated for matching calculation, and the operation is repeated until the matching degree meets the set threshold; In the step S3, the improved time sequence convolution network is used to predict the deformation offset of the current blanket, which specifically includes: The collected blanket tension, transmission speed, and roller shaft displacement are spliced into a feature matrix; The feature matrix is standardized; Multi-layer dilated causal convolution, with each layer having a kernel size of 3, is used for feature extraction. The feature extraction calculation method is as follows: ,in, Indicates the first The hidden state of the layer at time step sj. Representing the The weight matrix of the layer, For dilated convolution operations, Representing the The expansion factor of the layer Represented as the first Layer in time step The hidden state at that time is used as the input for the current convolution operation. Representing the Layer bias terms; Residual connection is made to the output of each layer of dilated causal convolution, and the calculation method is as follows: when the number of layers is greater than or equal to 2, wherein is the feature output of two layers before; The aggregated features are obtained by weighting and aggregating the features of different layers through an attention mechanism ; The features after the aggregation are input into independent prediction heads, and translation offset and rotation angle offset are output respectively wherein , , wherein, represents an independent fully connected layer.​ 2. The deep learning-based blanket embossing positioning method according to claim 1, wherein, In the step S2, the extraction method of the positioning anchor points specifically includes image preprocessing, key point detection, and anchor point screening; Image preprocessing: grayscale and Gaussian filter processing are performed on the continuous image sequence to suppress noise and enhance the contrast of the pattern edge; Key point detection: difference operation is performed on the continuous image sequence to identify the corner points and end points of the embossed edge and screen out the coarse positioning anchor point set; Anchor point screening: based on the spatial distribution density of the coarse positioning anchor points, the DBSCAN clustering algorithm is used to remove outliers and retain the clustering centers as the positioning anchor points.

3. The deep learning-based blanket embossing positioning method according to claim 2, characterized in that, The specific implementation of the key point detection includes feature enhancement, corner point detection, end point detection, and non-redundant screening; Feature enhancement: inter-frame difference is performed on 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, the gradient amplitude of the single frame image is calculated, and the feature enhancement graph is generated by weighted fusion of the difference image and the gradient amplitude; Corner point detection: the corner point response calculation is performed on the feature enhancement map to generate a corner point response matrix, and the calculation manner is as follows: wherein represents the corner point response matrix, M represents the gradient structure matrix, and k is an experience coefficient. The pixel points greater than the set threshold in the corner point response matrix are extracted as candidate corner points; Endpoint detection: edge detection is performed on the feature enhancement graph to generate a binary edge graph, and then all edge segments are extracted, the coordinates of the two end points of each edge segment are calibrated, the end points are defined as the pixel points with a degree of 1 in the edge segment, and the connectivity of the detected end points is verified: if the end points are connected to at least three consecutive edge pixels in the edge graph, they are determined as valid end points, otherwise they are regarded as noise and removed; Non-redundant screening: merging the candidate corner set and the effective endpoint set to generate a preliminary anchor point set, and performing non-maximum suppression on the preliminary anchor point set to obtain a non-redundant coarse positioning anchor point set.

4. The deep learning based blanket embossing positioning method according to claim 1, wherein, The specific implementation of obtaining the coarse positioning result in the ROI in the step S2 is as follows: First, the current frame ROI is aligned with the corresponding regions of the adjacent frames before and after by using a convolution network. Using the aligned image, the coordinates of the anchor points are calculated by using the centroid algorithm. Based on the coordinates of the anchor points, the geometric contour fitting of the embossed pattern edge is performed by using the least square method to obtain the coarse positioning result.

5. The deep learning based blanket embossing positioning method according to claim 1, wherein, In the offset prediction, the loss function of the translation offset adopts an L1 loss function, and the cosine similarity loss function is adopted for the rotation angle, and the expression of the loss function is: wherein, represents a predicted value, represents a target value.

6. The deep learning based blanket embossing positioning method according to claim 1, wherein, The specific process of the implementation of the spatial transformation correction of the coarse positioning result based on the predicted deformation offset to generate the positioning mask is as follows: First, the translation transformation matrix and the rotation transformation matrix are constructed by 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 result are transformed by using the composite transformation matrix to obtain the corrected coordinates. Based on the corrected coordinates, a closed contour region is drawn in the image, the pixel points in the closed contour region are assigned a value of 1, and the pixel points outside the region are assigned a value of 0 to generate the positioning mask.

Images