A UV plate layout pattern accurate matching method based on a twin network

Abstract

The application discloses a UV plate-making pattern accurate matching method based on a twin network, relates to the technical field of printing plate-making pattern matching, and comprises the following steps: receiving task information, and generating a plate-making matching constraint set; collecting bearing plate positioning information, establishing a plate-making space reference and a reference template set; solidifying a master plate image collected before transfer printing, and generating a plate-making observation record; inputting the reference template set and the plate-making observation record into an improved SiamFC model, performing light variation texture enhancement, multi-scale cross-correlation and cyclic consistency screening, and generating a pattern corresponding relation set; performing rigid pose solving and adjusting a master plate pose, re-collecting an image to generate a corrected pattern corresponding relation set; calculating a remaining matching residual error, correcting a plate-making alignment correction amount, and generating a plate-making pattern matching result. The application suppresses illumination and texture interference, drives pose correction by the remaining matching residual error, forms a closed loop of pattern matching, equipment alignment and result verification.

Description

Technical Field

[0001] This invention relates to the field of printing plate pattern matching technology, and in particular to a method for accurate matching of UV imposition patterns based on twin networks. Background Technology

[0002] UV imposition is a plate-making process that sequentially positions and transfers multiple unit patterns onto a substrate according to a pre-set layout. It is widely used in packaging printing, decorative printing, and fine pattern processing. During the imposition process, each unit pattern needs to maintain a corresponding position and orientation relationship with the target imposition unit on the substrate.

[0003] Existing UV panelizing equipment typically uses mechanical positioning, coordinate preset, or visual matching to align unit master plates. The visual matching method involves acquiring images of the unit master plates and comparing the acquired images with pre-stored reference patterns to obtain positional and angular deviations.

[0004] Existing visual matching methods mostly use grayscale correlation, template matching, or local feature point matching to determine the corresponding position of a pattern. However, due to the variability in lighting conditions, reflection status, and imaging position at the UV mosaicking site, the brightness distribution and local texture response in the acquired image may differ from the reference pattern.

[0005] When a pattern contains repeating textures, similar edges, or locally missing areas, single-scale comparison or unidirectional feature matching is prone to producing incorrect correspondences, causing the calculated positional and angular deviations to deviate from the actual alignment.

[0006] In addition, some existing methods adjust the pose of the master plate of the control device after completing one deviation calculation, without re-matching and verifying the adjusted pattern position.

[0007] Due to the combined effects of bearing plate positioning deviation, visual measurement deviation, and equipment execution deviation, residual translational and rotational deviations may still exist after a single adjustment. Furthermore, existing methods are unable to continue correcting based on the actual pattern state after adjustment, resulting in a lack of continuous verification relationship between pattern matching results and equipment alignment results.

[0008] Therefore, how to provide a method for accurate matching of UV mosaic patterns based on twin networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0009] One objective of this invention is to propose a precise matching method for UV pattern mosaicking based on Siamese networks. This invention fully utilizes Siamese network visual matching, spatial calibration, and rigid pose solving techniques, and describes in detail the implementation process of establishing a spatial reference for mosaicking, matching the reference pattern with the observed pattern, correcting the pose of the unit master pattern, and cyclically verifying the remaining matching residuals.

[0010] This invention has the advantages of strong resistance to light and similar texture interference, stable pattern correspondence, high equipment alignment accuracy, and closed-loop controllable correction process.

[0011] A method for accurate matching of UV tile patterns based on twin networks according to an embodiment of the present invention includes the following steps: Step 1: Receive the task information for precise matching of the pattern layout, perform matching object parsing, basic data association and matching requirement regularization, and generate a pattern layout matching constraint set; Step 2: Collect the positioning information of the bearing plate according to the panel matching constraint set, perform coordinate association and target unit mapping, and generate the panel space reference; Step 3: Perform target cell parsing and datum pattern binding on the panel matching constraint set, and generate a panel datum template set by combining the panel space datum; Step 4: According to the imposition space reference, control the UV imposition equipment to move the current unit master to the visual acquisition range of the target unit, acquire the image of the current unit master before curing and transfer, perform region extraction and spatial binding, and generate imposition observation records; Step 5: Input the mosaic baseline template set and mosaic observation records into the improved SiamFC model, and generate a pattern correspondence set through optical texture enhancement, multi-scale cross-correlation and cycle consistency filtering; Step 6: Perform rigid pose solving based on the pattern correspondence set and the panel space reference to generate panel alignment correction values; Step 7: Control the UV panelization equipment to adjust the pose of the current unit master according to the panelization alignment correction amount. After it is in place, re-acquire the image of the current unit master, perform region extraction and spatial binding, and generate a panelization correction observation record. Step 8: Input the mosaic benchmark template set and the mosaic calibration observation records into the improved SiamFC model, and generate the calibration pattern correspondence set through multi-scale cross-correlation and cyclic consistency screening; Step 9: Calculate the remaining matching residuals based on the set of correspondences of the calibration patterns, perform verification according to the set of panel matching constraints, and if the verification fails, correct the panel alignment correction amount based on the remaining matching residuals and re-perform the pose adjustment to generate the panel pattern matching result.

[0012] Optionally, step one specifically includes: Read the target panel unit identifier, current unit master plate identifier, and carrier plate identifier from the task information for accurate matching of the panel pattern; Perform integrity and object uniqueness checks on each identifier, establish task binding relationships between the target panel unit, the current unit master, and the carrier board, and generate panel matching object information; Based on the imposition matching object information, associate the imposition design file, the standard acquisition image of the unit master, and the spatial calibration relationship corresponding to the current UV imposition equipment; Verify the binding relationship between each basic data and the corresponding matching object, generate the basic data for puzzle matching, and terminate the current task if any matching object has not established a unique binding relationship or any basic data is missing; Read the matching residual threshold and loop count limit from the task information of precise matching of the pattern; perform a measurement unit uniformity and valid range check on the matching residual threshold; perform a positive integer check on the loop count limit; and generate the pattern matching requirements. Establish associations between layout matching object information, layout matching basic data, and layout matching requirements to generate a layout matching constraint set.

[0013] Optionally, step two specifically includes: Read the layout matching object information and layout matching basic data from the layout matching constraint set, determine the target layout unit and the support plate based on the layout matching object information, and call the layout design file and spatial calibration relationship based on the layout matching basic data; Collect the carrier plate positioning information bound to the carrier plate, verify the binding relationship and value range between the carrier plate positioning information and the carrier plate, and terminate the current task if the verification fails. Read the layout pose and target area boundary of the target layout unit from the layout design file, perform coordinate mapping on the target layout unit, and map the layout pose from the layout design coordinate system to the carrier plate coordinate system; Combine the carrier plate positioning information with the UV panelization equipment coordinate system to obtain the target pose of the target panelization unit, and transform the target area boundary to the UV panelization equipment coordinate system according to the same coordinate mapping. Establish the coordinate relationship between the UV panelization equipment coordinate system and the image coordinate system according to the spatial calibration relationship, map the transformed target area boundary to the image coordinate system, generate the image projection range of the target panelization unit, and associate the equipment target pose, image projection range and coordinate relationship to generate the panelization space reference.

[0014] Optionally, step three specifically includes: Read the imposition matching object information and imposition matching basic data from the imposition matching constraint set, locate the target unit design area in the imposition design file according to the target imposition unit identifier in the imposition matching object information, and call the unit master standard acquisition image bound to the current unit master according to the current unit master identifier; Analyze the target unit design area, read the target area boundary and target pattern direction, verify the consistency between the target area boundary and the target unit design area, and terminate the current task if the verification fails. According to the binding relationship between the target imposition unit and the current unit master in the imposition matching object information, the standard acquisition image of the unit master is bound as the reference pattern of the target imposition unit, and the execution direction of the standard acquisition image of the unit master is unified according to the direction of the target pattern. Based on the mosaic space reference, the boundary of the target region is mapped to the image coordinate system to generate the reference local position; According to the target area boundary after mapping, the size of the standard image of the unit master after orientation unification is performed, a reference template image is generated, and the reference template image is associated with the reference local position to generate a mosaic reference template set.

[0015] Optionally, step four specifically includes: Read the imposition matching object information from the imposition matching constraint set, determine the current unit master based on the imposition matching object information, and read the device target pose, image projection range and coordinate association from the imposition space reference; Based on the target pose of the equipment, a pose adjustment command is generated to control the alignment mechanism of the UV panelizing equipment to perform translation and rotation along the coordinate system of the UV panelizing equipment, so that the current unit master plate is moved to the visual acquisition range of the target unit. After the alignment mechanism stops moving, the positioning status is read. If the positioning verification fails, the current task is terminated. After the in-place verification is completed, the current position and orientation of the current unit master are maintained, and the vision acquisition device is triggered to acquire the image of the current unit master before the curing and transfer process is started. Bind the current cell master image to the current cell master, and verify whether the current cell master image completely covers the image projection range. If it does not completely cover the image projection range, terminate the current task. Determine the cropping boundary in the current unit master image according to the image projection range, crop the observation pattern image along the cropping boundary, record the cropping position of the cropping boundary in the image coordinate system, and bind the observation pattern image and the cropping position to the mosaic space reference according to the coordinate association to generate the mosaic observation record.

[0016] Optionally, the improved SiamFC model includes a dual-path isomorphic coding layer, an optically variable texture enhancement layer, a multi-scale cross-correlation layer, a cyclic consistency filtering layer, and a correspondence output layer. The optically variable texture enhancement layer embeds a structural steady-state enhancement mechanism, and the cyclic consistency filtering layer embeds a bidirectional backtracking mutual verification mechanism. The dual-path isomorphic coding layer uses a dual-branch network with shared weights to perform shallow, medium, and deep coding on the reference template image and the observation pattern image, generating reference multi-scale features and observation multi-scale features. The optically variable texture enhancement layer performs channel normalization and local texture response extraction on the baseline multi-scale features and the observed multi-scale features; The structural steady-state enhancement mechanism generates illumination suppression weights based on local brightness differences and the continuity of directional gradients; The illumination suppression weights are applied to the local texture response and superimposed with the corresponding multi-scale feature residuals to generate the baseline steady-state features and the observation steady-state features. The multi-scale cross-correlation layer performs channel-by-channel cross-correlation on the baseline steady-state features and observed steady-state features of the same scale; After unifying the resolution of the correlation responses at each scale, the correlation responses are stitched together and mapped to generate a fused correlation response. Based on this, candidate observation feature positions are determined for each benchmark feature position, and candidate correspondences are generated. The cyclic consistent filtering layer establishes candidate location neighborhood relationships based on candidate correspondence relationships; The two-way backtracking and mutual verification mechanism performs forward matching and reverse backtracking, generating positive correlation responses, reverse correlation responses, and backtracking positions; Candidate correspondences that do not return to the original reference feature unit are filtered out based on the backtracking position, and candidate correspondences with inconsistent direction or distance changes are filtered out based on the neighborhood relationship of the candidate position, generating valid correspondences; The correspondence output layer maps the effective correspondence back to the reference template image and the observation pattern image respectively, and generates the reference correspondence position and the observation correspondence position by combining the reference local position and the cropping position. Sigmoid normalization is performed on the positive and negative correlation responses respectively, and the smaller value in the normalization result is taken as the corresponding confidence level. The corresponding positions of the baseline, the corresponding positions of the observations, and the corresponding confidence levels are associated to generate a pattern correspondence set.

[0017] Optionally, step six specifically includes: The reference correspondence position, observation correspondence position, and correspondence confidence level are read from the pattern correspondence set. For correspondences where the reference correspondence position is repeated, the observation correspondence position is repeated, or both are repeated, the correspondence with the highest confidence level is retained. The current task will be terminated if any of the following situations occur: the number of retained correspondences is insufficient to meet the requirements of two-dimensional rigid pose solving, the reference corresponding positions are collinear, or the observation corresponding positions are collinear. If the above situation does not occur, divide the retained corresponding confidence level by the sum of the corresponding confidence levels to generate the corresponding solution weights; Taking the mapping of the observed corresponding position to the benchmark corresponding position as the solution direction, the weighted center of the benchmark corresponding position and the observed corresponding position is calculated according to the corresponding solution weights. The two sets of corresponding positions are centered and a weighted cross-covariance matrix is ​​constructed according to the corresponding solution weights. Singular value decomposition is performed on the weighted cross-covariance matrix, reflection correction is performed based on the determinant sign of the decomposition matrix, a local rotation matrix is ​​generated, and a local translation vector is generated based on the difference between the weighted center at the observation location after transformation by the local rotation matrix and the weighted center at the reference location. Read the coordinate association from the imposition space reference, and convert the local rotation matrix and local translation vector into rotation correction components and translation correction components in the UV imposition equipment coordinate system according to the coordinate association. Associate the rotation correction components and translation correction components to generate the imposition alignment correction amount.

[0018] Optionally, step seven specifically includes: Read the translation and rotation correction components from the imposition alignment correction amount, generate the pose adjustment command according to the translation and rotation directions of the UV imposition equipment coordinate system, and control the alignment mechanism of the UV imposition equipment to adjust the position and posture of the current unit master plate. After the alignment mechanism completes translation and rotation, it reads the positioning status fed back by the UV panelizing equipment. If the positioning status is not fed back, the vision acquisition device is not triggered. After the positioning status is fed back, the corrected pose of the current unit master is maintained. Before starting the curing and transfer process, re-acquire the current unit master image, read the image projection range from the imposition space reference, and verify the coverage integrity of the current unit master image over the image projection range. If the verification fails, terminate the current task. Determine the clipping boundary according to the image projection range, extract the correction observation pattern image from the current unit master image, and record the correction clipping position of the clipping boundary in the image coordinate system; The coordinate association is read from the mosaic spatial reference, and the correction observation pattern image and correction clipping position are bound to the mosaic spatial reference to generate the mosaic correction observation record.

[0019] Optionally, step eight specifically includes: Read the reference template image and reference local position from the mosaic reference template set, read the correction observation pattern image and correction clipping position from the mosaic correction observation record, and perform image normalization according to the reference branch input size and observation branch input size of the improved SiamFC model respectively; The normalized baseline template image and the corrected observation pattern image are input into the improved SiamFC model to generate baseline steady-state features and corrected observation steady-state features; Perform channel-wise cross-correlation on the baseline steady-state features and the corrected observation steady-state features of the same scale, stitch and map the correlation responses of each scale after unifying the resolution, generate the corrected fusion correlation response, determine the corrected observation feature position corresponding to the baseline feature position based on the corrected fusion correlation response, and generate the corrected candidate correspondence; Based on the candidate correspondence, establish the neighborhood relationship of the candidate position, perform forward matching from the benchmark to the correction observation and reverse backtracking from the correction observation to the benchmark on the candidate correspondence, and generate the forward correlation response, the reverse correlation response and the backtracking position; Delete the correction candidate correspondences where the backtracking position does not fall into the feature unit to which the corresponding reference feature position belongs, and delete the correction candidate correspondences where the direction and distance of adjacent positions are inconsistent according to the neighborhood relationship of the candidate positions, and generate the corrected valid correspondences; The benchmark feature positions and corrected observation feature positions in the corrected effective correspondence are mapped back to the benchmark template image and the corrected observation pattern image, respectively. The benchmark corresponding position and the observation corresponding position are generated by combining the benchmark local position and the corrected cropping position. Sigmoid normalization is performed on the positive and negative correlation responses respectively, and the smaller value in the normalization result is taken as the corresponding confidence level. The baseline corresponding position, the observation corresponding position and the corresponding confidence level are correlated to generate a set of calibration pattern correspondences.

[0020] Optionally, step nine specifically includes: The reference corresponding position, observation corresponding position, and corresponding confidence level are read from the set of calibration pattern correspondences. For correspondences where the reference corresponding position and observation corresponding position are duplicated, the one with the highest confidence level is retained. If any of the following situations occur: the number of retained correspondences is insufficient to meet the requirements of two-dimensional rigid pose solving, the reference corresponding positions are collinear, or the observation corresponding positions are collinear, the matching state will be set to abnormal, the mosaic pattern matching result will be generated, and the current task will be terminated. If the above situation does not occur, divide the retained corresponding confidence level by the sum of the corresponding confidence levels to generate residual solution weights; The solution direction is to map the observed position to the corresponding reference position. Two-dimensional rigid pose solution is performed according to the residual solution weight, combined with the coordinate association in the mosaic space reference. The solution results are converted into residual translation and residual rotation components in the coordinate system of the UV panelizing equipment. The magnitude of the residual translation component and the absolute value of the residual rotation component are calculated to generate the residual matching residual. Read the matching residual threshold and the maximum number of iterations from the layout matching requirements, and parse the matching residual threshold into translation residual threshold and rotation residual threshold; Establish a cyclic correction count with an initial value of zero. When the magnitude of the remaining translation component is not greater than the translation residual threshold and the absolute value of the remaining rotation component is not greater than the rotation residual threshold, set the matching state to pass. If the matching verification fails and the cycle correction count is less than the upper limit of the number of cycles, write the remaining translation component and the remaining rotation component into the panel alignment correction amount, and increment the cycle correction count by one. Based on the updated mosaic alignment correction, re-execute pose adjustment, image acquisition, region extraction, and spatial binding, and update the mosaic correction observation record; Input the mosaic reference template set and the updated mosaic correction observation record into the improved SiamFC model, update the correction pattern correspondence set, and perform two-dimensional rigid pose solution and matching verification again. When the cyclic correction count reaches the upper limit of the number of cycles and fails the matching verification, the matching status is set to failed. After the matching status is determined, the remaining matching residuals are associated with the matching status to generate the matching result of the mosaic pattern.

[0021] The beneficial effects of this invention are: This invention establishes a spatial reference for the mosaicking unit, unifying and binding the device target pose, image projection range, and coordinates of the target mosaicking unit. It also associates the reference template image, the observed pattern image, and their local positions with the same spatial reference, so that the positional relationship between the reference pattern and the real-time observed pattern has a unified coordinate basis, providing a continuous data foundation for pattern correspondence mapping and rigid pose solving.

[0022] Meanwhile, the improved SiamFC model incorporates a structural steady-state enhancement mechanism and a bidirectional backtracking mutual verification mechanism. Illumination suppression weights are generated by the continuity of local brightness difference and directional gradient, which weakens the impact of illumination changes and local reflections on texture response. Furthermore, forward matching, backward backtracking, and neighborhood relationship filtering are used to eliminate correspondences that cannot return to the original reference position or have inconsistent local changes, thereby enhancing the stability of pattern correspondences under conditions of repetitive textures, similar edges, and missing local patterns.

[0023] Furthermore, a weighted rigid pose solution is performed based on the pattern correspondence, converting the rotation and translation results in the image coordinate system into the panel alignment correction amount in the UV panelization equipment coordinate system. After the pose adjustment, the image is re-acquired, and the pattern matching and residual verification are re-executed, so that the visual matching result can directly drive the equipment alignment. The remaining deviation is continuously corrected based on the actual pattern state after correction, forming a closed-loop processing process of pattern matching, pose adjustment and result verification. Attached Figure Description

[0024] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a UV pattern matching method based on twin networks proposed in this invention. Figure 2 This is a schematic diagram of the structure of the improved SiamFC model proposed in this invention; Figure 3 This is a flowchart of the cyclic correction process based on residual matching, as proposed in this invention. Detailed Implementation

[0025] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0026] refer to Figures 1-3 A method for accurate matching of UV tile patterns based on twin networks includes the following steps: Step 1: Receive the task information for precise matching of the pattern layout, perform matching object parsing, basic data association and matching requirement regularization, and generate a pattern layout matching constraint set; Step 2: Collect the positioning information of the bearing plate according to the panel matching constraint set, perform coordinate association and target unit mapping, and generate the panel space reference; Step 3: Perform target cell parsing and datum pattern binding on the panel matching constraint set, and generate a panel datum template set by combining the panel space datum; Step 4: According to the imposition space reference, control the UV imposition equipment to move the current unit master to the visual acquisition range of the target unit, acquire the image of the current unit master before curing and transfer, perform region extraction and spatial binding, and generate imposition observation records; Step 5: Input the mosaic baseline template set and mosaic observation records into the improved SiamFC model, and generate a pattern correspondence set through optical texture enhancement, multi-scale cross-correlation and cycle consistency filtering; Step 6: Perform rigid pose solving based on the pattern correspondence set and the panel space reference to generate panel alignment correction values; Step 7: Control the UV panelization equipment to adjust the pose of the current unit master according to the panelization alignment correction amount. After it is in place, re-acquire the image of the current unit master, perform region extraction and spatial binding, and generate a panelization correction observation record. Step 8: Input the mosaic benchmark template set and the mosaic calibration observation records into the improved SiamFC model, and generate the calibration pattern correspondence set through multi-scale cross-correlation and cyclic consistency screening; Step 9: Calculate the remaining matching residuals based on the set of correspondences of the calibration patterns, perform verification according to the set of panel matching constraints, and if the verification fails, correct the panel alignment correction amount based on the remaining matching residuals and re-perform the pose adjustment to generate the panel pattern matching result.

[0027] In this embodiment, step one specifically includes: Receive the task information for accurate matching of the pattern, read the target pattern unit identifier, the current unit master plate identifier and the carrier plate identifier, perform integrity verification and object uniqueness verification on the three identifiers, terminate the current task when any identifier is missing, the same identifier corresponds to multiple objects or the three identifiers cannot establish a unique task binding relationship, associate the target pattern unit, the current unit master plate and the carrier plate, and generate pattern matching object information; Based on the imposition matching object information, the imposition design file is called according to the target imposition unit identifier, the standard acquisition image of the unit master is called according to the current unit master identifier, and the spatial calibration relationship is called according to the current UV imposition equipment. The imposition design file is checked to see if it contains the target imposition unit, if the standard acquisition image of the unit master is bound to the current unit master, and if the spatial calibration relationship corresponds to the current UV imposition equipment. If any basic data is missing or has not been uniquely bound to the corresponding matching object, the current task is terminated. The three basic data are associated to generate imposition matching basic data. Read the matching residual threshold and the upper limit of the number of loops. The matching residual threshold includes the translation residual threshold and the rotation residual threshold. Unify the translation residual threshold to the length unit used by the coordinate system of the UV panelization equipment and unify the rotation residual threshold to the angle unit used by the coordinate system of the UV panelization equipment. Check whether the two residual thresholds are positive values ​​and whether the upper limit of the number of loops is a positive integer. If the check fails, terminate the current task. If the check passes, generate the panelization matching requirements. Establish associations between the imposition matching object information, imposition matching basic data, and imposition matching requirements according to the current task, and generate an imposition matching constraint set; By unifying the matching objects, basic data, and verification conditions, consistent task constraints are provided for the collection of carrier board positioning information, the establishment of the panel space benchmark, pattern matching, and cyclic correction.

[0028] In this embodiment, step two specifically involves: Read the layout matching object information and layout matching basic data from the layout matching constraint set, determine the target layout unit and the support plate according to the target layout unit identifier and the support plate identifier, and call the layout design file and spatial calibration relationship; The carrier plate positioning information bound to the carrier plate is collected. The carrier plate positioning information is used to characterize the translation and rotation relationship of the carrier plate coordinate system relative to the UV panelization equipment coordinate system. The carrier plate identification is checked to see if it is consistent with the current task binding relationship. The translation and rotation relationship is compared with the effective positioning range recorded in the spatial calibration relationship. The current task is terminated when any value exceeds the effective positioning range. Read the layout pose and target area boundary of the target layout unit from the layout design file. Based on the origin correspondence and axial correspondence between the layout design coordinate system and the carrier plate coordinate system, map the layout pose to the carrier plate coordinate system. Then, according to the translation and rotation relationships in the carrier plate positioning information, convert it to the UV layout equipment coordinate system to generate the equipment target pose. Finally, convert the target area boundary to the UV layout equipment coordinate system according to the same coordinate mapping. Establish a coordinate association between the UV panelization equipment coordinate system and the image coordinate system according to the spatial calibration relationship. Map the transformed target area boundary to the image coordinate system according to the coordinate association to generate the image projection range of the target panelization unit. Terminate the current task when the mapping result exceeds the effective projection range of the coordinate association or the image projection range is empty. The device target pose, image projection range, and coordinates are associated according to the target panel unit to generate a panel space reference. By unifying the carrier plate positioning information, device target pose, image projection range, and coordinate association, a consistent spatial basis is provided for the generation of the panel reference template set and the visual acquisition of the current unit master.

[0029] In this embodiment, step three specifically includes: Read the imposition matching object information and imposition matching basic data from the imposition matching constraint set, locate the target unit design area in the imposition design file according to the target imposition unit identifier, call the unit master standard acquisition image bound to the current unit master according to the current unit master identifier, and terminate the current task when it cannot be uniquely located or called. Analyze the target unit design area, read the target area boundary and target pattern direction in vertex order, and verify whether the target area boundary is closed and whether the target pattern direction is valid. If the verification fails, terminate the current task. Read the target pose and coordinate association of the device from the panel space reference. Based on the rotation relationship and coordinate association in the target pose, convert the target pattern direction into the image target direction in the image coordinate system. Bind the unit master standard acquisition image as the reference pattern of the target panel unit. Take the horizontal axis of the unit master standard acquisition image as the master reference direction. Calculate the direction difference between the master reference direction and the image target direction. Rotate the unit master standard acquisition image according to the direction difference to generate the unit master standard acquisition image after the direction is unified. Based on the target pose of the device, the boundary of the target area is transformed to the coordinate system of the UV panelizing device, and then projected to the image coordinate system according to the coordinate association. When the boundary vertices exceed the effective projection range, the current task is terminated, and the projected boundary vertices are connected in the original order to generate the reference local position. The target size is determined based on the horizontal and vertical ranges of the reference local position. Bilinear interpolation is performed on the standard acquisition image of the unit master after the orientation is unified so that the width and height of the output image are consistent with the target size, and a reference template image is generated. The reference template image is associated with the reference local position to generate a panel reference template set. By unifying the orientation, size, and image coordinates of the reference pattern, a template basis is provided for improving the SiamFC model to call the reference template image and reflect the corresponding position of the pattern.

[0030] In this embodiment, step four specifically includes: Read the information of the imposition matching object from the imposition matching constraint set, determine the current unit master based on the current unit master identifier, read the device target pose, image projection range and coordinate association from the imposition space reference, and terminate the current task if any object cannot be read or is inconsistent with the binding relationship of the current task. Calculate translational and rotational deviations based on the target pose of the equipment and the current pose fed back by the UV panelizing equipment. Generate pose adjustment commands based on translational and rotational deviations. Verify whether the pose adjustment commands are within the effective movement range of the UV panelizing equipment. If they exceed the effective movement range, terminate the current task. Control the alignment mechanism to adjust the position and posture of the current unit master. After the alignment mechanism stops moving, read the positioning status. If no positioning status is fed back, terminate the current task. After the positioning verification is completed, the alignment mechanism is stopped and the position and posture of the current unit master are maintained. Before the curing and transfer process is started, the vision acquisition device is triggered to acquire the image of the current unit master and bind the image of the current unit master with the current unit master. If the image acquisition fails or the binding relationship is inconsistent, the current task is terminated. Check whether the current unit master image completely covers the image projection range. If it does not completely cover the range, terminate the current task. Determine the clipping boundary according to the image projection range. Extract the observation pattern image from the current unit master image along the clipping boundary and record the clipping position of the clipping boundary in the image coordinate system. Based on coordinate association, the observed pattern image and the cropping position are bound to the mosaic space reference to generate a mosaic observation record; By associating the observed pattern image and the cropping position with the tile space reference, observational evidence is provided to improve the matching of the reference pattern and the observed pattern in the SiamFC model.

[0031] In this embodiment, the improved SiamFC model includes a dual-path isomorphic coding layer, an optically variable texture enhancement layer, a multi-scale cross-correlation layer, a cyclic consistency filtering layer, and a correspondence output layer. The optically variable texture enhancement layer embeds a structural steady-state enhancement mechanism, and the cyclic consistency filtering layer embeds a bidirectional backtracking mutual verification mechanism. Specifically: Read the reference template image and reference local position from the mosaic reference template set, read the observation pattern image and cropping position from the mosaic observation record, normalize the reference template image into a 127×127×3 reference branch input tensor, normalize the observation pattern image into a 255×255×3 observation branch input tensor, maintain the aspect ratio of the original image during normalization, fill the missing edges with the mean edge pixel value of the corresponding image, divide the pixel value of each color channel by the maximum allowed pixel value of the corresponding color channel, and normalize to [0, 1]; The input tensors of the baseline branch and the observation branch are input into a dual-path isomorphic coding layer. The two coding branches share convolution parameters. The first-level convolution uses an 11×11 kernel, a 2-pixel stride, and 96 output channels. The second-level convolution uses a 5×5 kernel, a 1-pixel stride, and 256 output channels. The third to fifth levels of convolution all use a 3×3 kernel and a 1-pixel stride, with 384, 384, and 256 output channels, respectively. Batch normalization and ReLU activation are performed after the first four levels of convolution. Max pooling with a 3×3 pooling window and a 2-pixel stride is performed after the first and second levels of convolution, respectively. The fifth-level convolution directly outputs features. The results of the second, fourth, and fifth levels of convolution are read as shallow, medium, and deep features, respectively, to generate baseline multi-scale features and observation multi-scale features. The baseline features and observation features at the same scale maintain the same number of channels. The current task is terminated when any feature tensor is empty. The baseline multi-scale features and observed multi-scale features are input into the optical texture enhancement layer. The spatial mean and standard deviation are calculated for each scale by channel. The feature value is subtracted from the corresponding spatial mean and then divided by the sum of the corresponding standard deviation and 10^-6 to complete channel normalization. The normalized result is subjected to 3×3 depth convolution, with one 3×3 convolution kernel configured for each channel to generate local texture response. The structural steady-state enhancement mechanism calculates the local brightness difference and the continuity of the directional gradient at each feature location. The local brightness difference is the absolute value of the difference between the mean of the current channel and the mean of the 3×3 neighboring channels. The continuity of the directional gradient is the mean of the cosine similarity between the current gradient vector and the gradient vector of 8 neighbors after non-negative truncation. When calculating the cosine similarity, 10^-6 is added to the product of the magnitudes of the gradient vectors. Local brightness difference and directional gradient continuity are normalized according to the minimum and maximum values ​​in the feature map of the same scale. When the range is 0, the corresponding normalization result is set to 0. The local brightness difference normalization result is subtracted from 1 and averaged with the directional gradient continuity normalization result to generate the illumination suppression weight. The illumination suppression weights are applied to the local texture response at each position, and residual superposition and ReLU activation are performed with the corresponding multi-scale features to generate the baseline steady-state features and the observation steady-state features. The baseline steady-state features and observed steady-state features of the same scale are cross-correlated channel by channel, and the correlation responses at each scale are generated by summing along the channel dimension. The mid-level and deep correlation responses are unified to the spatial resolution of the shallow correlation response by bilinear interpolation. The three scale correlation responses are concatenated along the scale dimension and mapped to a single-channel fused correlation response by 1×1 convolution. The local maxima are filtered for each baseline feature position by 3×3. The three observed feature positions are retained from high to low according to the fused correlation response to generate candidate correspondences. Based on the candidate correspondence, an 8-neighborhood candidate position neighborhood relationship is established. The fusion correlation response from the baseline feature position to the candidate observation feature position is recorded as the positive correlation response. The bidirectional backtracking mutual verification mechanism uses the observation steady-state feature corresponding to the candidate observation feature position as the reverse correlation kernel. Cross-correlation is performed in the baseline steady-state feature. The position with the largest reverse correlation response is taken as the backtracking position. When the backtracking position does not return to the original baseline feature unit, the corresponding candidate correspondence is deleted. For the retained candidate correspondences, calculate the displacement direction and displacement distance. Within the neighborhood of the candidate position, calculate the median and median absolute deviation of the displacement direction deviation and displacement distance deviation, respectively. If any deviation exceeds the sum of the corresponding median and 2.5 times the median absolute deviation, delete the corresponding candidate correspondence. If the median absolute deviation is 0, only retain the candidate correspondences with deviations equal to the corresponding median, and generate valid correspondences. The correspondence output layer maps the baseline feature positions and observed feature positions in the effective correspondence back to the baseline template image and observed pattern image respectively according to the ratio of feature resolution to input image resolution. It generates baseline correspondence positions and observed correspondence positions by combining the baseline local position and the cropping position. It performs Sigmoid normalization on the positive correlation response and the negative correlation response respectively, and takes the smaller value of the two as the correspondence confidence level. It associates the baseline correspondence position, the observed correspondence position and the correspondence confidence level to generate a pattern correspondence set. Both the original SiamFC model and the improved SiamFC model adopt a dual-branch coding structure with shared weights and use the cross-correlation response between the baseline feature and the observed feature to determine the corresponding position. The similarity between the two is that they retain the basic structure of dual-path isomorphic coding and feature cross-correlation. The difference is that the improved SiamFC model adds an optical texture enhancement layer and a structural steady-state enhancement mechanism after the shared coding branch, changes the original single-scale cross-correlation to multi-scale cross-correlation, and adds a cyclic consistency screening layer and a two-way backtracking mutual verification mechanism before the correlation response output. Then, the correspondence output layer completes the position backtracking and correspondence confidence calculation, so that the original model outputs a single correlation peak position judgment method to transform the pattern correspondence generation method with illumination suppression, two-way verification, neighborhood consistency screening and spatial position backtracking. By enhancing optical texture, multi-scale cross-correlation, cyclic consistency screening, and bidirectional backtracking verification, the panel reference template set and panel observation records are converted into a pattern correspondence set, providing the reference corresponding position, observation corresponding position, and correspondence confidence level for the rigid pose solution of the panel alignment correction quantity.

[0032] In this embodiment, step six specifically includes: The baseline correspondence position, observation correspondence position, and correspondence confidence level are read from the pattern correspondence set. First, the correspondences are grouped according to the baseline correspondence position, and the correspondence with the highest confidence level is retained in each group. Then, the retained correspondences are grouped according to the observation correspondence position, and the correspondence with the highest confidence level is retained in each group. The number of correspondences after retention is counted. Mean centering is performed on the baseline correspondence position and the observation correspondence position respectively. The baseline-centered coordinate matrix and the observation-centered coordinate matrix are constructed. Singular value decomposition is performed on the two centered coordinate matrices respectively. If the number of correspondences is less than 3 and the ratio of the second singular value to the first singular value of any centered coordinate matrix is ​​less than 10^-6, it is determined that there are insufficient correspondences or the corresponding positions are collinear and the current task is terminated. Calculate the sum of the corresponding confidence levels after retention. Terminate the current task when the sum of the corresponding confidence levels is 0. Divide each corresponding confidence level by the sum of the corresponding confidence levels to generate the corresponding solution weight. According to the corresponding solution weights, the horizontal and vertical coordinates of the reference corresponding position and the observation corresponding position are weighted and summed respectively to generate the weighted center of the reference corresponding position and the weighted center of the observation corresponding position. The weighted center of the reference corresponding position is subtracted from each reference corresponding position, and the weighted center of the observation corresponding position is subtracted from each observation corresponding position to generate two sets of centered corresponding positions. Taking the mapping of the observed corresponding position to the corresponding reference position as the solution direction, the outer product of each centralized observed corresponding position and the corresponding centralized reference corresponding position is performed, and the weighted sum is calculated according to the corresponding solution weights to generate a weighted cross-covariance matrix; Perform singular value decomposition on the weighted cross-covariance matrix, multiply the right singular matrix by the transpose of the left singular matrix to generate a local rotation matrix. When the determinant of the local rotation matrix is ​​less than 0, invert the sign of the last column of the right singular matrix and recalculate the local rotation matrix to eliminate the reflection transformation. The local rotation matrix is ​​used to transform the weighted center at the corresponding observation position. The difference between the weighted center at the corresponding reference position and the transformation result is calculated to generate a local translation vector. Read the coordinate association from the imposition space reference, and transform the local rotation matrix and local translation vector from the image coordinate system to the UV imposition device coordinate system according to the coordinate association. Generate rotation correction components and translation correction components, associate the rotation correction components and translation correction components, and generate imposition alignment correction amount. By weighting according to the degree of confidence, correcting for reflection, and transforming coordinates, the pattern correspondence set is converted into a set of imposition alignment correction quantities that the UV imposition equipment can perform, providing a basis for correction of the position and orientation adjustment of the current unit master.

[0033] In this embodiment, step seven specifically includes: Read the translation and rotation correction components from the alignment correction amount, generate pose adjustment instructions according to the translation and rotation directions of the UV panelization equipment coordinate system, and verify whether the translation and rotation correction components are within the effective motion range of the UV panelization equipment. Terminate the current task if any component exceeds the effective motion range. The pose adjustment command is sent to the alignment mechanism of the UV panelization equipment. The alignment mechanism is controlled to adjust the position and posture of the current unit master according to the translation correction component and the rotation correction component. After the alignment mechanism stops moving, the positioning status fed back by the UV panelization equipment is read. If no positioning status is fed back, the vision acquisition device is kept off and the current task is terminated. After the feedback status is in place, the alignment mechanism is stopped, the corrected pose of the current unit master is maintained, and the vision acquisition device is triggered to re-acquire the image of the current unit master before the curing and transfer process starts. If the image acquisition fails, the current task is terminated. Read the image projection range from the tile space reference, compare the image projection range with the effective imaging range of the current unit master image that has been re-acquired, and terminate the current task when the image projection range is not completely covered by the effective imaging range. Determine the cropping boundary according to the image projection range, extract the correction observation pattern image from the re-acquired current unit master image along the cropping boundary, and record the correction cropping position of the cropping boundary in the image coordinate system. Read the coordinate association from the mosaic spatial reference, and bind the correction observation pattern image and correction clipping position to the mosaic spatial reference according to the coordinate association to generate the mosaic correction observation record; By maintaining the corrected pose and re-performing image acquisition, region extraction, and spatial binding, a corrected observation basis consistent with the mosaic spatial reference is provided for generating the correction pattern correspondence set.

[0034] In this embodiment, step eight specifically includes: Read the reference template image and reference local position from the mosaic reference template set, read the correction observation pattern image and correction clipping position from the mosaic correction observation record, and perform size unification, edge filling and pixel value normalization according to the reference branch input size and observation branch input size of the improved SiamFC model respectively. Terminate the current task when any image is missing or the normalization result is empty. The normalized baseline template image and the corrected observation pattern image are input into the improved SiamFC model. Multi-scale feature extraction is performed by the dual-path isomorphic coding layer, channel normalization and local texture response extraction are performed by the optical texture enhancement layer, and illumination suppression weights are calculated by the structural steady-state enhancement mechanism and residual superposition is performed to generate baseline steady-state features and corrected observation steady-state features. The multi-scale cross-correlation layer performs channel-wise cross-correlation on the baseline steady-state features and the corrected observation steady-state features at the same scale. The correlation responses at each scale are unified at a uniform resolution and then stitched together to generate the corrected fusion correlation response. The corrected observation feature positions are determined based on the corrected fusion correlation response as the position of each baseline feature, and the corrected candidate correspondence is generated. The cyclic consistent filtering layer performs forward matching and reverse backtracking on the correction candidate correspondences, generates forward correlation response, reverse correlation response and backtracking position, deletes correction candidate correspondences whose backtracking position does not return to the original reference feature unit, and deletes correction candidate correspondences with inconsistent displacement direction or displacement distance changes according to the neighborhood relationship of the candidate position, and generates valid correction correspondences. The correspondence output layer maps the baseline feature positions and corrected observation feature positions in the corrected effective correspondence back to the baseline template image and the corrected observation pattern image, respectively. It combines the baseline local position and the corrected cropping position to generate the baseline correspondence position and the observation correspondence position. The smaller value in the normalized results of the positive correlation response and the negative correlation response is used as the correspondence confidence level. The baseline correspondence position, the observation correspondence position and the correspondence confidence level are associated to generate the corrected pattern correspondence set. By re-performing pattern matching after the pose adjustment of the current unit master, the corrected actual pattern state is converted into a set of corrected pattern correspondences, providing the corrected corresponding position basis for the calculation of the remaining matching residuals and the matching verification.

[0035] In this embodiment, step nine specifically includes: The reference correspondence position, observation correspondence position, and correspondence confidence level are read from the calibration pattern correspondence set. First, the correspondences are grouped according to the reference correspondence position, and the correspondence with the highest confidence level is retained in each group. Then, the retained correspondences are grouped according to the observation correspondence position, and the correspondence with the highest confidence level is retained in each group. The number of correspondences after retention is counted. Mean centering and singular value decomposition are performed on the baseline correspondence position and the observation correspondence position respectively. If the number of correspondences is less than 3 and the ratio of the second singular value to the first singular value of any centered coordinate matrix is ​​less than 10^-6, the matching status is set to abnormal, the mosaic pattern matching result is generated and the current task is terminated. When there is no insufficient correspondence or collinearity of corresponding positions, calculate the sum of the corresponding confidence levels after retention. When the sum of the corresponding confidence levels is 0, set the matching status to abnormal, generate the matching result of the mosaic pattern and terminate the current task. When the sum of the corresponding confidence levels is greater than 0, divide each corresponding confidence level by the sum of the corresponding confidence levels to generate residual solution weights. Taking the mapping of the observed position to the reference position as the solution direction, two-dimensional rigid pose solution is performed according to the residual solution weight. The coordinate association is read from the panel space reference. The local rotation matrix and local translation vector obtained by the solution are converted into the residual rotation component and residual translation component in the coordinate system of the UV panel equipment. The absolute value of the residual rotation component and the magnitude of the residual translation component are calculated to generate the residual matching residual. Read the matching residual threshold and the upper limit of the number of loops from the imposition matching requirements, parse the matching residual threshold into translation residual threshold and rotation residual threshold, and establish a loop correction count with an initial value of 0; If the magnitude of the remaining translation component is not greater than the translation residual threshold and the absolute value of the remaining rotation component is not greater than the rotation residual threshold, the matching status is set to pass. If the matching verification fails and the cycle correction count is less than the maximum number of cycles, write the remaining translation component into the translation correction component in the panel alignment correction quantity, write the remaining rotation component into the rotation correction component in the panel alignment correction quantity, and increment the cycle correction count by 1. Adjust the position and orientation of the current unit master according to the updated panel alignment correction amount. After it is in place, re-acquire the image of the current unit master, perform region extraction and spatial binding, update the panel correction observation record, input the panel reference template set and the updated panel correction observation record into the improved SiamFC model, update the correction pattern correspondence set, perform correspondence deduplication and two-dimensional rigid pose solution again, update the remaining matching residuals and perform matching verification. When the cyclic correction count reaches the upper limit of the number of cycles and fails the matching verification, the matching status is set to failed. After the matching status is determined, the remaining matching residuals are associated with the matching status to generate the matching result of the mosaic pattern. By writing the remaining translation and rotation components into the panel alignment correction amount and cyclically triggering pose adjustment, image acquisition, pattern matching, and result verification, the panel pattern matching result reflects the actual alignment state of the current unit master after correction.

[0036] Example 1: To verify the feasibility of the present invention in practice, the present invention was applied to the unit master alignment scenario of a UV imposition equipment. 240 target imposition units were selected as verification samples and divided into four imaging states: normal lighting, uneven local brightness, repetitive texture, and missing local patterns. 60 target imposition units were set for each state and distinguished by anonymous sample numbers.

[0037] The translation residual threshold is set to 0.05mm, the rotation residual threshold is set to 0.03°, and the maximum number of cycles is set to 3. The initial lateral translation deviation of the current unit master is distributed between -1.20mm and 1.20mm, the initial longitudinal translation deviation is distributed between -1.00mm and 1.00mm, and the initial rotation deviation is distributed between -1.20° and 1.20°.

[0038] According to the method of this invention, a set of panel matching constraints, a panel spatial reference, a set of panel reference templates, and panel observation records are generated. The set of panel reference templates and panel observation records are input into the improved SiamFC model to generate a pattern correspondence set. Based on the pattern correspondence set, a two-dimensional rigid pose solution is performed to generate the panel alignment correction amount and adjust the pose of the current unit master plate. After the pose adjustment, the current unit master plate image is re-acquired to generate a correction pattern correspondence set. The remaining matching residuals are calculated and cyclic correction is performed to finally generate the panel pattern matching result.

[0039] A comparison was made between the traditional local feature matching scheme, the original SiamFC scheme, and the scheme of this invention. The three schemes use the same target panel unit, initial pose deviation, two-dimensional rigid pose solution, device pose adjustment, matching residual threshold, and cyclic correction process.

[0040] Traditional local feature matching schemes use local feature description and nearest neighbor correspondence to generate correspondences. The original SiamFC scheme uses shared weight dual-branch encoding and single-scale cross-correlation to generate correspondences. The scheme of this invention uses an improved SiamFC model to generate correspondences.

[0041] The reference positions and observation positions of each verification sample, which are pre-labeled, are used as the statistical reference for the correspondence. The remaining translation and rotation components obtained by solving the two-dimensional rigid pose of the correction pattern correspondence set after one pose adjustment are used as the basis for the deviation statistics. The effective correspondence accuracy, the mismatch rate of repeated texture samples, the average remaining translation deviation after one correction, the average remaining rotation deviation after one correction, the final matching pass rate, and the average number of pose corrections are statistically analyzed. The average number of pose corrections includes the pose adjustment performed based on the initial alignment correction amount and the cyclic correction performed after failing the matching verification. The specific comparison results are shown in Table 1.

[0042] Table 1. Comparison of Alignment Effects of Different Pattern Matching Schemes

[0043] As shown in Table 1, the effective correspondence accuracy of the proposed solution is 98.8%, which is 9.6 percentage points and 3.7 percentage points higher than that of the traditional local feature matching solution and the original SiamFC solution, respectively. In 60 repeated texture samples, the traditional local feature matching solution, the original SiamFC solution, and the proposed solution had 9, 5, and 1 mismatched samples, respectively, with mismatch rates of 15.0%, 8.3%, and 1.7%.

[0044] The average residual translation deviation after one correction in this invention is 0.024 mm, and the average residual rotation deviation is 0.015°, both of which are lower than the corresponding matching residual threshold. 239 out of 240 target mosaic units passed the matching verification, and the final matching pass rate was 99.6%. One sample that failed corresponds to a correction observation pattern image with a large local pattern missing range. The matching status was set to failed after the cyclic correction count reached the upper limit of the number of cycles.

[0045] Compared with the traditional local feature matching scheme and the original SiamFC scheme, the average pose correction times of the present invention are reduced to 1.09 times, indicating that most target panel units meet the matching residual threshold after the first pose adjustment. A small number of samples that fail the correction are corrected again after the remaining translation and rotation components are written into the panel alignment correction amount.

[0046] The above results show that the structural steady-state enhancement mechanism reduces the impact of local brightness changes on texture response, multi-scale cross-correlation preserves pattern structure information at different spatial scales, bidirectional backtracking and mutual verification mechanism and candidate position neighborhood relationship screening reduce erroneous correspondences caused by repeated textures, and residual residual cyclic verification allows the residual translation and rotation components after the first pose adjustment to continue to be written into the panel alignment correction amount. As can be seen from the above implementation results, the present invention can complete UV panel pattern matching, current unit master pose adjustment and correction result verification.

[0047] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for precise matching of UV tile patterns based on twin networks, characterized in that, Includes the following steps: Step 1: Receive the task information for precise matching of the pattern layout, perform matching object parsing, basic data association and matching requirement regularization, and generate a pattern layout matching constraint set; Step 2: Collect the positioning information of the bearing plate according to the panel matching constraint set, perform coordinate association and target unit mapping, and generate the panel space reference; Step 3: Perform target cell parsing and datum pattern binding on the panel matching constraint set, and generate a panel datum template set by combining the panel space datum; Step 4: According to the imposition space reference, control the UV imposition equipment to move the current unit master to the visual acquisition range of the target unit, acquire the image of the current unit master before curing and transfer, perform region extraction and spatial binding, and generate imposition observation records; Step 5: Input the mosaic baseline template set and mosaic observation records into the improved SiamFC model, and generate a pattern correspondence set through optical texture enhancement, multi-scale cross-correlation and cycle consistency filtering; Step 6: Perform rigid pose solving based on the pattern correspondence set and the panel space reference to generate panel alignment correction values; Step 7: Control the UV panelization equipment to adjust the pose of the current unit master according to the panelization alignment correction amount. After it is in place, re-acquire the image of the current unit master, perform region extraction and spatial binding, and generate a panelization correction observation record. Step 8: Input the mosaic benchmark template set and the mosaic calibration observation records into the improved SiamFC model, and generate the calibration pattern correspondence set through multi-scale cross-correlation and cyclic consistency screening; Step 9: Calculate the remaining matching residuals based on the corresponding relationship set of the calibration patterns, perform verification according to the panel matching constraint set, and if the verification fails, correct the panel alignment correction amount based on the remaining matching residuals and re-perform the pose adjustment to generate the panel pattern matching result.

2. The method for precise matching of UV mosaic patterns based on twin networks according to claim 1, characterized in that, Step one specifically involves: Read the target panel unit identifier, current unit master plate identifier, and carrier plate identifier from the task information for accurate matching of the panel pattern; Perform integrity and object uniqueness checks on each identifier, establish task binding relationships between the target panel unit, the current unit master, and the carrier board, and generate panel matching object information; Based on the imposition matching object information, associate the imposition design file, the standard acquisition image of the unit master, and the spatial calibration relationship corresponding to the current UV imposition equipment; Verify the binding relationship between each basic data and the corresponding matching object, generate the basic data for puzzle matching, and terminate the current task if any matching object has not established a unique binding relationship or any basic data is missing; Read the matching residual threshold and loop count limit from the task information of precise matching of the pattern; perform a measurement unit uniformity and valid range check on the matching residual threshold; perform a positive integer check on the loop count limit; and generate the pattern matching requirements. Establish associations between layout matching object information, layout matching basic data, and layout matching requirements to generate a layout matching constraint set.

3. The method for precise matching of UV mosaic patterns based on twin networks according to claim 1, characterized in that, Step two specifically involves: Read the layout matching object information and layout matching basic data from the layout matching constraint set, determine the target layout unit and the support plate based on the layout matching object information, and call the layout design file and spatial calibration relationship based on the layout matching basic data; Collect the carrier plate positioning information bound to the carrier plate, verify the binding relationship and value range between the carrier plate positioning information and the carrier plate, and terminate the current task if the verification fails. Read the layout pose and target area boundary of the target layout unit from the layout design file, perform coordinate mapping on the target layout unit, and map the layout pose from the layout design coordinate system to the carrier plate coordinate system; Combine the carrier plate positioning information with the UV panelization equipment coordinate system to obtain the target pose of the target panelization unit, and transform the target area boundary to the UV panelization equipment coordinate system according to the same coordinate mapping. Establish the coordinate relationship between the UV panelization equipment coordinate system and the image coordinate system according to the spatial calibration relationship, map the transformed target area boundary to the image coordinate system, generate the image projection range of the target panelization unit, and associate the equipment target pose, image projection range and coordinate relationship to generate the panelization space reference.

4. The method for precise matching of UV mosaic patterns based on twin networks according to claim 1, characterized in that, Step three specifically involves: Read the imposition matching object information and imposition matching basic data from the imposition matching constraint set, locate the target unit design area in the imposition design file according to the target imposition unit identifier in the imposition matching object information, and call the unit master standard acquisition image bound to the current unit master according to the current unit master identifier; Analyze the target unit design area, read the target area boundary and target pattern direction, verify the consistency between the target area boundary and the target unit design area, and terminate the current task if the verification fails. According to the binding relationship between the target imposition unit and the current unit master in the imposition matching object information, the standard acquisition image of the unit master is bound as the reference pattern of the target imposition unit, and the execution direction of the standard acquisition image of the unit master is unified according to the direction of the target pattern. Based on the mosaic space reference, the boundary of the target region is mapped to the image coordinate system to generate the reference local position; According to the target area boundary after mapping, the size of the standard image of the unit master after orientation unification is performed, a reference template image is generated, and the reference template image is associated with the reference local position to generate a mosaic reference template set.

5. The method for precise matching of UV mosaic patterns based on twin networks according to claim 1, characterized in that, Step four specifically involves: Read the imposition matching object information from the imposition matching constraint set, determine the current unit master based on the imposition matching object information, and read the device target pose, image projection range and coordinate association from the imposition space reference; Based on the target pose of the equipment, a pose adjustment command is generated to control the alignment mechanism of the UV panelizing equipment to perform translation and rotation along the coordinate system of the UV panelizing equipment, so that the current unit master plate is moved to the visual acquisition range of the target unit. After the alignment mechanism stops moving, the positioning status is read. If the positioning verification fails, the current task is terminated. After the in-place verification is completed, the current position and orientation of the current unit master are maintained, and the vision acquisition device is triggered to acquire the image of the current unit master before the curing and transfer process is started. Bind the current cell master image to the current cell master, and verify whether the current cell master image completely covers the image projection range. If it does not completely cover the image projection range, terminate the current task. Determine the cropping boundary in the current unit master image according to the image projection range, crop the observation pattern image along the cropping boundary, record the cropping position of the cropping boundary in the image coordinate system, and bind the observation pattern image and the cropping position to the mosaic space reference according to the coordinate association to generate the mosaic observation record.

6. The method for precise matching of UV mosaic patterns based on twin networks according to claim 1, characterized in that, The improved SiamFC model includes a dual-path isomorphic coding layer, an optically variable texture enhancement layer, a multi-scale cross-correlation layer, a cyclic consistency screening layer, and a correspondence output layer. The optically variable texture enhancement layer embeds a structural steady-state enhancement mechanism, and the cyclic consistency screening layer embeds a bidirectional backtracking mutual verification mechanism. The dual-path isomorphic coding layer uses a dual-branch network with shared weights to perform shallow, medium, and deep coding on the reference template image and the observation pattern image, generating reference multi-scale features and observation multi-scale features. The optically variable texture enhancement layer performs channel normalization and local texture response extraction on the baseline multi-scale features and the observed multi-scale features; The structural steady-state enhancement mechanism generates illumination suppression weights based on local brightness differences and the continuity of directional gradients; The illumination suppression weights are applied to the local texture response and superimposed with the corresponding multi-scale feature residuals to generate the baseline steady-state features and the observation steady-state features. The multi-scale cross-correlation layer performs channel-by-channel cross-correlation on the baseline steady-state features and observed steady-state features of the same scale; After unifying the resolution of the correlation responses at each scale, the correlation responses are stitched together and mapped to generate a fused correlation response. Based on this, candidate observation feature positions are determined for each benchmark feature position, and candidate correspondences are generated. The cyclic consistent filtering layer establishes candidate location neighborhood relationships based on candidate correspondence relationships; The two-way backtracking and mutual verification mechanism performs forward matching and reverse backtracking, generating positive correlation responses, reverse correlation responses, and backtracking positions; Candidate correspondences that do not return to the original reference feature unit are filtered out based on the backtracking position, and candidate correspondences with inconsistent direction or distance changes are filtered out based on the neighborhood relationship of the candidate position, generating valid correspondences; The correspondence output layer maps the effective correspondence back to the reference template image and the observation pattern image respectively, and generates the reference correspondence position and the observation correspondence position by combining the reference local position and the cropping position. Sigmoid normalization is performed on the positive and negative correlation responses respectively, and the smaller value in the normalization result is taken as the corresponding confidence level. The corresponding positions of the baseline, the corresponding positions of the observations, and the corresponding confidence levels are associated to generate a pattern correspondence set.

7. The method for accurate matching of UV mosaic patterns based on twin networks according to claim 1, characterized in that, Step six specifically involves: The reference correspondence position, observation correspondence position, and correspondence confidence level are read from the pattern correspondence set. For correspondences where the reference correspondence position is repeated, the observation correspondence position is repeated, or both are repeated, the correspondence with the highest confidence level is retained. The current task will be terminated if any of the following situations occur: the number of retained correspondences is insufficient to meet the requirements of two-dimensional rigid pose solving, the reference corresponding positions are collinear, or the observation corresponding positions are collinear. If the above situation does not occur, divide the retained corresponding confidence level by the sum of the corresponding confidence levels to generate the corresponding solution weights; Taking the mapping of the observed corresponding position to the benchmark corresponding position as the solution direction, the weighted center of the benchmark corresponding position and the observed corresponding position is calculated according to the corresponding solution weights. The two sets of corresponding positions are centered and a weighted cross-covariance matrix is ​​constructed according to the corresponding solution weights. Singular value decomposition is performed on the weighted cross-covariance matrix, reflection correction is performed based on the determinant sign of the decomposition matrix, a local rotation matrix is ​​generated, and a local translation vector is generated based on the difference between the weighted center at the observation location after transformation by the local rotation matrix and the weighted center at the reference location. Read the coordinate association from the imposition space reference, and convert the local rotation matrix and local translation vector into rotation correction components and translation correction components in the UV imposition equipment coordinate system according to the coordinate association. Associate the rotation correction components and translation correction components to generate the imposition alignment correction amount.

8. The method for accurate matching of UV mosaic patterns based on twin networks according to claim 1, characterized in that, Step seven specifically involves: Read the translation and rotation correction components from the imposition alignment correction amount, generate the pose adjustment command according to the translation and rotation directions of the UV imposition equipment coordinate system, and control the alignment mechanism of the UV imposition equipment to adjust the position and posture of the current unit master plate. After the alignment mechanism completes translation and rotation, it reads the positioning status fed back by the UV panelizing equipment. If the positioning status is not fed back, the vision acquisition device is not triggered. After the positioning status is fed back, the corrected pose of the current unit master is maintained. Before starting the curing and transfer process, re-acquire the current unit master image, read the image projection range from the imposition space reference, and verify the coverage integrity of the current unit master image over the image projection range. If the verification fails, terminate the current task. Determine the clipping boundary according to the image projection range, extract the correction observation pattern image from the current unit master image, and record the correction clipping position of the clipping boundary in the image coordinate system; The coordinate association is read from the mosaic spatial reference, and the correction observation pattern image and correction clipping position are bound to the mosaic spatial reference to generate the mosaic correction observation record.

9. The method for accurate matching of UV mosaic patterns based on twin networks according to claim 1, characterized in that, Step eight specifically involves: Read the reference template image and reference local position from the mosaic reference template set, read the correction observation pattern image and correction clipping position from the mosaic correction observation record, and perform image normalization according to the reference branch input size and observation branch input size of the improved SiamFC model respectively; The normalized baseline template image and the corrected observation pattern image are input into the improved SiamFC model to generate baseline steady-state features and corrected observation steady-state features; Perform channel-wise cross-correlation on the baseline steady-state features and the corrected observation steady-state features of the same scale, stitch and map the correlation responses of each scale after unifying the resolution, generate the corrected fusion correlation response, determine the corrected observation feature position corresponding to the baseline feature position based on the corrected fusion correlation response, and generate the corrected candidate correspondence; Based on the candidate correspondence, establish the neighborhood relationship of the candidate position, perform forward matching from the benchmark to the correction observation and reverse backtracking from the correction observation to the benchmark on the candidate correspondence, and generate the forward correlation response, the reverse correlation response and the backtracking position; Delete the correction candidate correspondences where the backtracking position does not fall into the feature unit to which the corresponding reference feature position belongs, and delete the correction candidate correspondences where the direction and distance of adjacent positions are inconsistent according to the neighborhood relationship of the candidate positions, and generate the corrected valid correspondences; The benchmark feature positions and corrected observation feature positions in the corrected effective correspondence are mapped back to the benchmark template image and the corrected observation pattern image, respectively. The benchmark corresponding position and the observation corresponding position are generated by combining the benchmark local position and the corrected cropping position. Sigmoid normalization is performed on the positive and negative correlation responses respectively, and the smaller value in the normalization result is taken as the corresponding confidence level. The baseline corresponding position, the observation corresponding position and the corresponding confidence level are correlated to generate a set of calibration pattern correspondences.

10. The method for precise matching of UV mosaic patterns based on twin networks according to claim 1, characterized in that, Step nine specifically involves: The reference corresponding position, observation corresponding position, and corresponding confidence level are read from the set of calibration pattern correspondences. For correspondences where the reference corresponding position and observation corresponding position are duplicated, the one with the highest confidence level is retained. If any of the following situations occur: the number of retained correspondences is insufficient to meet the requirements of two-dimensional rigid pose solving, the reference corresponding positions are collinear, or the observation corresponding positions are collinear, the matching state will be set to abnormal, the mosaic pattern matching result will be generated, and the current task will be terminated. If the above situation does not occur, divide the retained corresponding confidence level by the sum of the corresponding confidence levels to generate residual solution weights; The solution direction is to map the observed position to the corresponding reference position. Two-dimensional rigid pose solution is performed according to the residual solution weight, combined with the coordinate association in the mosaic space reference. The solution results are converted into residual translation and residual rotation components in the coordinate system of the UV panelizing equipment. The magnitude of the residual translation component and the absolute value of the residual rotation component are calculated to generate the residual matching residual. Read the matching residual threshold and the maximum number of iterations from the layout matching requirements, and parse the matching residual threshold into translation residual threshold and rotation residual threshold; Establish a cyclic correction count with an initial value of zero. When the magnitude of the remaining translation component is not greater than the translation residual threshold and the absolute value of the remaining rotation component is not greater than the rotation residual threshold, set the matching state to pass. If the matching verification fails and the cycle correction count is less than the upper limit of the number of cycles, write the remaining translation component and the remaining rotation component into the panel alignment correction amount, and increment the cycle correction count by one. Based on the updated mosaic alignment correction, re-execute pose adjustment, image acquisition, region extraction, and spatial binding, and update the mosaic correction observation record; Input the mosaic reference template set and the updated mosaic correction observation record into the improved SiamFC model, update the correction pattern correspondence set, and perform two-dimensional rigid pose solution and matching verification again. When the cyclic correction count reaches the upper limit of the number of cycles and fails the matching verification, the matching status is set to failed. After the matching status is determined, the remaining matching residuals are associated with the matching status to generate the matching result of the mosaic pattern.

Images