A Deep Learning-Based Method for Determining Print Color Registration Deviation

By generating a set of color edge anchor points and constructing a color edge relationship graph, deep learning technology is used to extract local offset evidence of color separation contours from printed images, solving the misjudgment problem of existing color registration deviation judgment methods and achieving more accurate color registration deviation judgment.

CN123118534APending Publication Date: 2026-11-17DONGGUAN SHENGDE IND INVESTMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611329214.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for determining color registration deviation in printing based on image block classification cannot effectively confirm whether detected color edges or ghosting are formed by the relative displacement of color separation layers. Furthermore, they are prone to misjudgment when ink layer concentration, illumination reflection, and local printing defects produce color separation appearances similar to color registration deviations, and lack constraints at the geometric level.

Method used

By generating a set of color edge anchor points, a deep feature extraction network is used to extract local offset evidence of color separation contours from the detection area, a color edge relationship graph is constructed, and local offset evidence of adjacent color edge anchor points is aggregated through a graph neural network to generate the color registration deviation distribution of the page and output the color registration deviation judgment result.

Benefits of technology

It reduces misjudgments caused by ink layer variations and texture interference, lowers the probability of missing true color registration deviations in weak contour areas, and improves the accuracy and reliability of color registration deviation determination.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118534A_ABST
    Figure CN123118534A_ABST
Patent Text Reader

Abstract

This invention relates to the fields of printed image detection, deep learning, and graph data processing, specifically a method for determining color registration deviation in printing based on deep learning. The method acquires a pre-press color separation layer and the image to be inspected. It generates color edge anchor points and locates the detection area based on the overlapping and adjacency relationships of the color separation contours. Local offset vectors and evidence credibility are extracted through a dual-branch approach combining color response and contour geometry. A color edge relationship graph is constructed according to the continuity of the same-color contours, spatial adjacency, and cross-color common edge relationships. A graph neural network is used to aggregate adjacent anchor point evidence, and the node status is checked using a bidirectional mapping closed-loop residual to form the color registration deviation distribution on the printing plate. The method outputs the offset color separation, direction, range, and type. This invention, through the synergistic constraint of geometric evidence and spatial relationships on the printing plate, suppresses color interference caused by ink layer changes, illumination reflection, and substrate texture, reduces the propagation of isolated anomalies, and improves the consistency of color registration deviation identification, location, and type differentiation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of printed image detection, deep learning, and image data processing, specifically a method for determining printing color registration deviation based on deep learning. Background Technology

[0002] Color registration deviation determination for color printed materials uses the relative positions of multiple color separation layers on the substrate as the detection object. Existing implementations typically set up area scan cameras, line scan cameras, and synchronous illumination modules at the receiving end of the printing press or the quality inspection station. The detection area is located based on registration marks, paper edges, or preset page coordinates. The continuously acquired printed images are cropped into several image blocks. After distortion correction, scale unification, brightness equalization, noise suppression, and color representation transformation, the image blocks, together with pre-labeled normal registration samples, horizontal offset samples, vertical offset samples, and samples of different deviation levels, constitute training data. A convolutional neural network extracts color edges, ghosting, repeated contours, and local color separation features through multi-layer convolution and downsampling. Then, the classification layer outputs the category probability of each color registration state. The detection end completes the determination of a single image block based on the category probability, and forms the determination conclusion of the entire printed material by the majority results of multiple detection areas or the maximum abnormal probability. At the same time, the corresponding image block, determination category, and acquisition position are saved for subsequent review and sample update.

[0003] During model training and online deployment, existing solutions typically establish a corresponding sample library around a fixed product layout. They obtain samples with different directions and degrees of color misregistration by manually adjusting the printing plate position, moving color separation layers, or screening abnormal printed sheets during production. Then, they expand the training images using rotation, cropping, brightness variations, blurring, and noise overlay. After model training, the system uses the layout positioning results to map the image to be inspected to the reference coordinates, calls the classification model block by block according to the pre-configured detection areas, and summarizes the category probability, number of abnormal areas, and continuous printed sheet judgment results for each image block. Some solutions also use qualified sample sheets or digital originals as reference images, performing local registration or feature similarity calculations between the reference image and the image to be inspected, and then inputting the difference map and convolutional features into the classification layer. When the product pattern, printing material, or printing batch changes, the model is retrained or adjusted by supplementing corresponding samples to match the classification boundary with the current product's color appearance and pattern structure.

[0004] Existing image patch classification methods still rely on local color appearance and texture statistics as the basis for judgment. There is a lack of pre-press color separation contour correspondence between different detection areas and the continuity constraint of the same color plate offset in the page space. Therefore, it is impossible to confirm from the geometric relationship level whether the detected color edges or ghosting are formed by the relative displacement of the color separation layer. When ink layer concentration, illumination reflection, substrate texture or local printing defects produce a color separation appearance similar to color registration deviation, independent image patches may give an abnormal category, while the true color plate offset may lack sufficient classification features in areas with weak local contours or little color overlap. When multiple image patches are summed using the majority result or the maximum anomaly probability, local misjudgment will directly affect the conclusion of the entire printed product. The essence is that there is a lack of a constraint mechanism that can incorporate color separation geometric evidence and page correlation evidence into the judgment process, which makes it easy to confuse true color registration displacement with non-geometric color anomalies. Summary of the Invention

[0005] The purpose of this invention is to provide a method for determining printing color registration deviation based on deep learning, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A deep learning-based method for determining color registration deviation in printing includes: acquiring a pre-press color separation layer and a print image to be inspected corresponding to the pre-press color separation layer; generating a set of color edge anchor points based on the contour overlap relationship of each pre-press color separation layer; and locating the detection area corresponding to each color edge anchor point in the print image to be inspected.

[0008] Local offset evidence of color separation contours is obtained from each detection region through a deep feature extraction network;

[0009] A color edge relationship graph is constructed based on the continuity relationship of the same color contour and the spatial adjacency relationship between the color edge anchor points. The local offset evidence of adjacent color edge anchor points is aggregated through a graph neural network to obtain the node offset state.

[0010] Generate the color registration deviation distribution of the page layout based on the offset status of each node and output the color registration deviation judgment result.

[0011] Preferably, the generation of the color edge anchor point set includes: extracting closed contours, intersecting contours and adjacent contours from each prepress color separation layer, projecting the contours of different prepress color separation layers onto a unified page coordinate system, and forming a color edge topology description according to the color separation identifier, contour intersection order, edge direction and adjacency relationship.

[0012] Select a layout area containing at least two color separation contours and having a definite intersection order from the color edge topology description, and record the center position of the layout area, the color separation identifiers, the edge direction sequence and the contour connection relationship as the color edge anchor point.

[0013] Preferably, the deep feature extraction network includes a color response branch and a contour geometry branch;

[0014] The color response branch converts the detection area into a color contrast representation and generates each color separation response map based on the color separation participation identifier of the corresponding color edge anchor point;

[0015] The contour geometry branch extracts gradient direction, edge phase and contour connection structure from the detection area, and performs cross-attention fusion on each color response map according to the correspondence between the participating color divisions;

[0016] The fused features are input into the offset regression layer to obtain the local offset vectors of each non-reference color segment relative to the reference color segment and the credibility of the offset evidence.

[0017] Preferably, the color edge relationship diagram uses the color edge anchor point as the graph node, and establishes graph edges based on the same color contour continuity relationship, the page space adjacency relationship, and the cross-color common edge relationship, respectively;

[0018] Write the local offset evidence and the color edge topology description into the graph node, and write the anchor point coordinate difference, the contour tangential correspondence and the color separation identifier into the graph edge;

[0019] The graph neural network performs corresponding message transformations for different graph edge relationships and aggregates the messages of adjacent graph nodes in the local coordinate system of the colored edge anchor point to obtain the node offset state, which includes the offset direction, offset continuity state, and evidence credibility.

[0020] Preferably, locating the detection area corresponding to each color edge anchor point includes: extracting candidate edges in the printed image to be inspected, and generating candidate topological descriptions according to the direction order, color response order, and intersection connection relationship of the candidate edges;

[0021] The candidate topology description is matched with the color edge topology description of the color edge anchor point, and a local coordinate system of anchor point composed of contour tangent and contour normal is established for the candidate edges that form the correspondence.

[0022] Based on the local coordinate system mapping detection area of ​​the anchor point, when there are multiple candidate correspondences, joint assignment is completed by combining the arrangement order of adjacent colored edge anchor points and the contour connection relationship.

[0023] Preferably, the training of the color response branch and the contour geometry branch includes: constructing a first sample pair with the same geometric position relationship but different color brightness or ink layer response, and a second sample pair with similar color appearance but relatively shifted color separation contours;

[0024] The color appearance representation and geometric alignment representation in the sample are encoded separately, and the color appearance representation is exchanged between the first sample pairs and recombined with the original geometric alignment representation;

[0025] The local offset vectors before and after the constraint exchange remain consistent, and the geometric alignment representation is constrained by the color-separation movement annotation of the second sample pair, so that the local offset evidence in the detection stage is jointly generated by the geometric alignment representation and the color edge topology description.

[0026] Preferably, the generation of the local offset evidence further includes: inputting the reference image region of the color edge anchor point and the corresponding detection region into the contour geometric branch with shared parameters in a forward order and a reverse order, respectively, to obtain a forward offset vector from the reference image region to the detection region and a reverse offset vector from the detection region to the reference image region.

[0027] The reference contour points are mapped sequentially using the forward offset vector and the reverse offset vector. The credibility of the offset evidence is generated based on the closed-loop residual between the mapped contour points and the original reference contour points and the bidirectional contour correspondence consistency. The credibility of the offset evidence and the forward offset vector are written together into the corresponding graph node.

[0028] Preferably, the graph neural network sets message transformation paths for the same-color contour continuity relationship, the layout space adjacency relationship, and the cross-color common edge relationship, respectively;

[0029] Based on the contour tangent and contour normal of the color edge anchor point, the offset vectors of adjacent graph nodes are transformed to the local coordinate system of the target graph node;

[0030] Based on the credibility of the offset evidence and the contour tangential correspondence in the graph edge, a message gating value is generated. The relational messages output by each message transformation path are aggregated, and the offset vector, offset continuity state, and credibility of the evidence of the target graph node are updated using the aggregation result.

[0031] Preferably, in the color edge relationship diagram, clockwise directed anchor point loops and counterclockwise directed anchor point loops are generated from the color edge anchor points on the same closed printing contour, and the node offset state of each diagram node is sequentially converted to the unified layout coordinate system along the two directions and propagated to the starting diagram node.

[0032] The message credibility of the corresponding graph edge is updated based on the difference in the closed state formed by the two propagation paths at the starting graph node. Then, the updated message credibility is used to perform secondary relationship aggregation on each graph node in the directed anchor loop, and the node offset state obtained by the secondary relationship aggregation is written into the page color deviation distribution.

[0033] Preferably, the generation of the color registration deviation distribution of the page layout includes: projecting the node offset state of each image node onto the unified page layout coordinate system according to the color separation identifier;

[0034] Separate common translation components, local residuals that vary continuously along the printing outline, and isolated residuals that lack graph connectivity from the node offset states corresponding to the same color separation.

[0035] An overall deviation submap is formed based on the common translation components, a regional gradual deviation submap is formed based on the local residuals, and the isolated residuals are recorded as evidence of discontinuous offset.

[0036] By combining the credibility of evidence from each graph node with the overall deviation subgraph and the regional gradient deviation subgraph, a color deviation determination result is output, which includes offset color separation, spatial range, offset direction, and deviation type.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] 1. By extracting closed contours, intersecting contours, and adjacent contours from the prepress color separation layer, and establishing color edge anchor points based on color separation identifiers, contour intersection order, edge direction, and adjacency relationship, the detection area in the printed image to be inspected obtains a geometric reference corresponding to the prepress color separation structure. The color response branch and the contour geometric branch respectively form a color separation response map and a contour connection representation. After fusion, the local offset vector of each non-reference color separation relative to the reference color separation is output, so that the color registration anomaly is transformed from an appearance category into geometric evidence with color separation object and offset direction. After constructing a color edge relationship map based on the continuous relationship of the same color contour, the adjacency relationship of the page space, and the common edge relationship across colors, the graph neural network can aggregate local offset evidence between associated anchor points, so that the offset state that appears continuously along the same color separation contour remains consistent in the page coordinates, and separates isolated color anomalies that lack the support of associated nodes from the continuous deviation distribution, reducing misjudgments caused by ink layer changes, local stains, and texture interference, and reducing the probability of missing the true color registration deviation in weak contour areas.

[0039] 2. The correspondence between color edge topology description and candidate topology description enables the detection area to be located based on the contour intersection order and connection structure. Local coordinates of anchor points are established through contour tangent and contour normal, reducing the impact of overall sheet translation, rotation, and local deformation on anchor point correspondence. Separate training of color appearance representation and geometric alignment representation ensures that local offset evidence remains less dependent on color brightness and ink layer response changes. Forward offset vectors, reverse offset vectors, and contour point closed-loop residuals together form the credibility of the evidence. The bidirectional propagation results of the directed anchor point loop further correct the credibility of the image edge message, reducing the continuous propagation of erroneous anchor point correspondences in the image. After the node offset state is projected onto the unified page coordinates according to the color separation identifier, common translation components, continuous local residuals, and isolated residuals can be separated to form an overall deviation sub-map and a regional gradient deviation sub-map, ensuring that the output includes the color separation where the offset occurred, the spatial range, the offset direction, and the deviation type. Attached Figure Description

[0040] Figure 1 Flowchart for overall judgment of color registration deviation in printing;

[0041] Figure 2 Flowchart for generating evidence of color-bordered anchor point positioning and local offset;

[0042] Figure 3 This is a graph showing the variation of closed-loop residuals with anchor points;

[0043] Figure 4 A flowchart for generating color edge relationship aggregation and page color registration deviation;

[0044] Figure 5 A comparison chart of offsets before and after relation aggregation;

[0045] Figure 6 This is a distribution diagram of relational message gating. Detailed Implementation

[0046] In one embodiment, please refer to Figure 1A deep learning-based method for determining color registration deviation in printing includes: acquiring pre-press color separation layers and the corresponding printed images to be inspected; generating a set of color edge anchor points based on the contour overlap relationship of each pre-press color separation layer; locating the detection areas corresponding to each color edge anchor point in the printed image to be inspected; obtaining local offset evidence of color separation contours from each detection area through a deep feature extraction network; constructing a color edge relationship graph based on the continuity relationship and spatial adjacency relationship of the same color contours between color edge anchor points; aggregating local offset evidence of adjacent color edge anchor points through a graph neural network to obtain node offset states; generating a color registration deviation distribution based on each node offset state and outputting the color registration deviation determination result. In this embodiment, the pre-press color separation layers use raster data that maintains consistency with the actual plate-making data in terms of plate size, layer order, and color separation identifiers. Each pre-press color separation layer records the ink application area of ​​the corresponding ink in a unified plate coordinate system. The acquisition identifier bound to the current printed sheet is written into the processing queue. The processor converts the printed image to be inspected into a unified page coordinate system based on the page positioning information. The color edge anchor points are determined by the overlapping, adjacent or converging structures formed by different color separation contours in the same page area. The detection area is cut around the color edge anchor points and retains the contour segments connected to adjacent anchor points. The depth feature extraction network converts the color response and contour geometric information in the detection area into a local offset vector relative to the reference color separation and the corresponding evidence credibility. The color edge relationship graph connects the color edge anchor points with the same color contour continuity relationship, page space adjacency relationship and cross-color common edge relationship. The graph neural network propagates and corrects the local offset evidence under the unified coordinate semantics. The page color registration deviation distribution records the overall offset, continuous local offset and non-continuous offset evidence according to the color separation identifier. The judgment result is generated after comparing the deviation distribution according to the color registration allowable range corresponding to the printing task.

[0047] In this embodiment, the correspondence between the prepress color separation layer and the image to be inspected is determined by the printing task identifier, the page version identifier, and the sheet acquisition identifier. When reading data, the processor checks the page range, pixel coordinate direction, and layer origin of each color separation layer. If the color separation layer comes from a vector page file, each graphic object is projected onto the corresponding raster layer according to its color separation attributes while preserving the object boundaries. If the color separation layer comes from a rasterized plate-making file, the ink value of each color separation channel is directly read and a color separation identifier mapping table is established. Before the image to be inspected enters the analysis, the overall translation, rotation, and scale changes of the entire sheet relative to the unified page coordinate system are estimated using the page outline, stable graphic corner points, or reserved positioning graphics. The processor performs an overall coordinate correction on the acquired image. This overall coordinate correction only eliminates the common pose changes caused by sheet transport and image acquisition, without changing the relative positions of each color separation within the printed image to be inspected. This preserves the color contour separation information corresponding to the color registration deviation. After the correction is completed, the detection area is read according to the range index of the color edge anchor point in the unified page coordinate system. Simultaneously, the pre-press color separation sub-image, color separation identifiers, and contour connection records corresponding to the detection area are read. The processor marks sheets with missing color separation layers, inconsistent page versions, or those for which overall coordinates cannot be established as having incomplete data and does not send them to the color registration deviation judgment network. This ensures that the data entering subsequent processing has a definite color separation source and page coordinate basis.

[0048] Specifically, the reference color separation is selected from the color separations participating in the current color edge anchor point. The selection criteria are the connectivity length of the corresponding color separation contour in the neighborhood of the anchor point, the number of contour intersections, and the continuous coverage relationship with adjacent color edge anchor points. The processor sorts each candidate color separation by structural stability and records the color separation with the highest ranking as the reference color separation for the color edge anchor point. The remaining color separations are recorded as non-reference color separations. The selection only determines the reference direction of the local offset vector without changing the content of the image to be inspected. After global coordinate correction, the detection area forms a color contrast representation, a gradient direction representation, and a pre-press color separation prior representation. The color contrast representation is composed of the relative differences between the acquired color channels and retains the ink overlay. The overall color change formed by printing is represented by the gradient direction, which consists of the local change direction and magnitude of each color channel. The pre-printing color separation prior representation consists of the inked area, outline position, and color separation identifier in the corresponding color separation sub-image. The deep feature extraction network encodes the above representations on the same spatial scale, establishes the correspondence between different color separation responses at the outline intersection position, and establishes the directional continuity relationship of the same color separation response at the outline extension position. Then, the offset regression layer outputs the local offset vector of each non-reference color separation relative to the reference color separation in the horizontal and vertical directions along the unified page. The transformation relationship between the local offset vector and the local coordinate system of the anchor point is preserved so that the coordinates between the graph nodes can be unified.

[0049] Furthermore, each color edge anchor point corresponds to a graph node in the color edge relationship graph. The initial state of the graph node includes the anchor point's layout coordinates, the color separation identifier, the reference color separation identifier, the local offset vector, the evidence credibility level, the contour tangent, the contour normal, and the color edge topological description. The graph edges store the layout coordinate difference, local coordinate transformation relationship, common color separation identifier, and contour direction correspondence information between the source node and the target node according to the relationship type. After reading the initial state of the graph node, the graph neural network generates corresponding adjacency messages for different relationship types, transforms the offset vector in the local coordinates of the adjacent nodes to the local coordinates of the target node, and then, based on the evidence credibility level of the source node and the graph... The proportion of edge contour direction information control messages entering the target node is controlled by the edge contour direction. After the target node gathers the same-color contour continuous messages, the page space adjacency messages, and the cross-color common edge messages, it updates its own offset vector and offset continuity state. The node state obtained by relation aggregation is restored to the unified page coordinate system and written into the page grid according to the color separation identifier. The continuity constraint from the relation graph is maintained between adjacent grids covered by the same color separation. Non-continuous markers are retained around local evidence without graph connection support. The processor generates a color separation deviation distribution based on the color separation and reads the offset color separation, offset direction, spatial range, and deviation type from the color separation deviation distribution as the judgment result.

[0050] In a preferred embodiment, the generation of the color edge anchor point set includes: extracting closed contours, intersecting contours, and adjacent contours from each prepress color separation layer; projecting the contours of different prepress color separation layers onto a unified page coordinate system; forming a color edge topology description according to color separation identifiers, contour intersection order, edge direction, and adjacency relationship; selecting page areas containing at least two color separation contours and having a definite intersection order from the color edge topology description; recording the center position of the page area, the color separation identifiers, the edge direction sequence, and the contour connection relationship as color edge anchor points; in practice, performing connected region tracing and extracting the inner and outer boundaries of the inked area of ​​each prepress color separation layer; recording the boundary that connects the beginning and end and encloses the inked area as a closed contour; and recording the boundary that has branch connections within the same color separation or with other color separation contours on the page. In planar coordinates, intersecting boundaries are denoted as cross contours. Structures where different color separation boundaries are approximately parallel and continuously spaced within a local area are denoted as adjacent contours. The processor sorts the boundary points along the contour direction and records the contour direction before and after entering the intersection position, the order of color separation, and the connection type between each contour at the intersection position, thereby generating a color edge topology description independent of color intensity. For regions with only a single straight edge and lacking other color separation contours nearby, no color edge anchor points are established. For regions containing multiple color separation intersections but with repeated intersection orders, their connection paths with adjacent contours are merged and recorded, so that the anchor point identity can be determined by both the local intersection structure and the extended connection structure. Finally, the color edge anchor points are indexed according to the layout coordinates, and the anchor points on the boundary of the same printed object are written into the continuous relationship table.

[0051] In a preferred embodiment, please refer to Figure 2 The process of locating the detection area corresponding to each color edge anchor point includes: extracting candidate edges in the printed image to be inspected; generating candidate topological descriptions according to the direction order, color response order, and intersection connection relationship of the candidate edges; performing topological correspondence between the candidate topological descriptions and the color edge topological descriptions of the color edge anchor points; establishing a local coordinate system for anchor points composed of contour tangent and contour normal for the candidate edges that form a correspondence; mapping the detection area according to the local coordinate system of anchor points; when multiple candidate correspondences exist, completing joint assignment by combining the arrangement order of adjacent color edge anchor points and contour connection relationship; specifically, calculating the color change direction within the anchor point search range determined by the overall coordinate correction and connecting candidate edges with continuous directions; encoding the candidate edges in the order of extension from the intersection center outwards; and determining the color response order based on the relative color change on both sides of the candidate edge and the printing... The previous color separation identifiers are jointly determined, and the intersection connection relationship is represented by the entry order and exit order of candidate edges at the intersection center. The processor performs a hierarchical comparison between the candidate topology description and the color edge topology description. First, it checks the number of participating color separations and the intersection type, then it checks the cyclic order of the edge direction, and then it checks the connection direction with the adjacent anchor points along the outer contour. Candidate structures that meet the corresponding relationship are written into the candidate assignment set. When there are multiple candidates with similar structures in the same area, the local similarity of a single anchor point is not used for direct selection. Instead, the candidate positions of the adjacent anchor points are read at the same time. Joint assignment is completed by maintaining the anchor point arrangement order, contour connection direction and page adjacency relationship. After selection, a local horizontal axis is established with the average tangent of the matching contour, and a local vertical axis is established with the normal pointing to the outside of the inked area. Based on this, the detection area with normalized direction is truncated.

[0052] Preferably, the deep feature extraction network includes a color response branch and a contour geometry branch. The color response branch converts the detection region into a color-opposite representation and generates each color-separation response map based on the color-separation identifiers of the corresponding color edge anchor points. The contour geometry branch extracts the gradient direction, edge phase, and contour connection structure from the detection region and performs cross-attention fusion on each color-separation response map according to the correspondence between the participating color separations. The fused features are input into the offset regression layer to obtain the local offset vectors of each non-reference color separation relative to the reference color separation and the credibility of the offset evidence. The color response branch uses a shared space coding layer to extract the comprehensive color features of the detection region, and then uses the color-separation identifiers as conditional information to perform color-separation projection on the comprehensive color features, so that each color-separation response map represents the current color separation in the detection region. The relative response at each position is calculated, and the contour geometry branch calculates multi-directional gradients for the color contrast representation and color separation response map respectively. The gradient extrema are connected to form candidate contour segments, and geometric features including edge phase, curvature change and connection order are obtained through convolutional encoding. Cross-attention fusion uses non-reference color separation geometric features as queries and reference color separation geometric features as keys and values ​​to establish corresponding weights between the same contour direction and adjacent contour positions, and applies the corresponding weights to the two color separation response maps to form color edge alignment features. The offset regression layer generates local offset vectors based on the difference in the centroid of the response on the local horizontal axis and local vertical axis of the color edge alignment features. At the same time, it generates the credibility of offset evidence based on the proportion of contour segments that can form stable correspondences, the consistency of corresponding directions, and the subsequent bidirectional mapping results.

[0053] In a preferred embodiment, training the color response branch and the contour geometry branch includes: constructing a first sample pair with the same geometric position relationship but different color brightness or ink layer response, and a second sample pair with similar color appearance but relatively shifted color separation contours; encoding the color appearance representation and geometric alignment representation in the samples respectively; exchanging the color appearance representation between the first sample pairs and recombining it with the original geometric alignment representation; constraining the local offset vectors before and after the exchange to remain consistent, and constraining the geometric alignment representation based on the color separation movement annotation of the second sample pair, so that the local offset evidence in the detection stage is jointly generated by the geometric alignment representation and the color edge topology description. The first sample pair is selected from qualified printed sheets at the same page position, or the color brightness, comprehensive color offset, and local offset are changed while keeping the pre-press color separation contour position unchanged. The part reflection and ink layer uniformity are generated. The second sample pair is obtained by performing a relative movement of the non-reference color separation response map in a known direction and re-synthesizing it with the unmoved reference color separation response map. During the training process, the samples are input into the shared space encoding layer. The color encoding path outputs the color appearance representation, and the geometric encoding path outputs the geometric alignment representation. After the processor swaps the color appearance representation of the first sample pair, it combines it with the original geometric alignment representation and reconstructs the color separation response. If the offset outputs before and after the swap are different, the color encoding path and the geometric encoding path are updated according to the output difference, so that the information read by the offset regression is focused on the contour position relationship. The known movement direction and movement amount of the second sample pair are used as the supervision label of the local offset vector. At the same time, the color edge topology describes the corresponding position of the constraint intersection order and contour connection relationship in the feature space.

[0054] Furthermore, the generation of local offset evidence also includes: inputting the reference image region and the corresponding detection region of the color edge anchor point into the contour geometric branch with shared parameters in a forward and reverse order, respectively, to obtain a forward offset vector from the reference image region to the detection region and a reverse offset vector from the detection region to the reference image region; mapping the reference contour points sequentially using the forward and reverse offset vectors, generating the offset evidence credibility level based on the closed-loop residual between the mapped contour points and the original reference contour points and the bidirectional contour correspondence consistency, and writing the offset evidence credibility level and the forward offset vector together into the corresponding graph node. In practice, a set of contour points is selected from the color-separated contours of the reference image region according to the contour extension order, and the forward offset is performed. The vector maps the contour point set to the detection area. The contour geometry branch reads the detected contour points that are adjacent to the mapped position and have the same direction in the detection area. The reverse offset vector then maps the detected contour points back to the reference image area, thus obtaining the closed-loop return position of each reference contour point. The bidirectional contour correspondence consistency is determined by the color segment identifier, edge direction and connection order pointed to by the forward and reverse matching. When the forward and reverse mappings fall on the corresponding contour segments of the same color and the connection order is consistent, it is recorded as a consistent correspondence. When any mapping crosses to other color contours or changes the intersection order, it is recorded as a conflicting correspondence. The closed-loop residual and the bidirectional contour correspondence consistency together constitute the credibility of the evidence used by the graph nodes when receiving and sending relational messages.

[0055] In this embodiment, the credibility of the offset evidence is specifically expressed by the following formula.

[0056]

[0057] ;

[0058] Where i represents the index of the color edge anchor point, and c represents the non-base color separation used for judgment relative to the base color separation. This represents the number of contour points selected from the reference contour of the i-th colored edge anchor point. This represents the original coordinates of the k-th reference contour point within the i-th colored edge anchor point. This represents the closed-loop return coordinates obtained after the reference contour points are sequentially mapped forward and backward for color separation c. This represents the diagonal length of the reference image region corresponding to the i colored edge anchor points. This represents the closed-loop residual after regional-scale normalization. This indicates the consistency of the bidirectional contours, and is normalized to between -1 and 1 according to the conformity of the contour direction, color coding, and connection order. The symbol represents the credibility of the offset evidence of color separation c at the i-th color edge anchor point. exp represents the Euclidean norm of the two-dimensional coordinate difference; for example, if four reference contour points are selected for a certain color edge anchor point, the diagonal length of the detection area is 10 image length units, and the sum of the closed-loop coordinate distances of the four contour points is 4 image length units, then the normalized closed-loop residual is 1 / 10. When the bidirectional contour correspondence consistency is 8 / 10, the confidence level of the offset evidence is about 0.81. If the closed-loop return position is further closer to the original reference contour point and the contour correspondence remains unchanged, the confidence level of the offset evidence obtained by substituting it increases accordingly.

[0059] Please refer to Figure 3 , Figure 3 The horizontal axis represents the anchor point sequence formed by the scanning of the page layout and the numbering of the color border anchor points, and the vertical axis represents the closed-loop residual after the scale of the detected area is normalized. The discrete marks on the curve correspond to the closed-loop return error of different anchor points in the same batch of printed sheets. Figure 3 Corresponding to the closed-loop residual symbol in the text Its value is derived from the reference contour point. Closed-loop return point after forward and reverse mapping After accumulating the two-dimensional distances between them, and then combining them with the number of contour points... and the diagonal length of the region We obtain it by normalization. Figure 3 The overall low residual range of the main curve indicates that the candidate topology, contour direction, and color separation correspondence of most color edge anchor points can achieve stable closure. When a few anchor points show local elevation, it indicates that the contour return position at that point has deviated more significantly from the original reference contour point. This usually requires verification in conjunction with whether there are occlusions, scratches, ink shortages, duplicate textures, or mismatched intersection sequences in the detection area. The mean reference line is used to observe the average level of the closed-loop error of the entire page, while the upper observation line is used to highlight relatively prominent abnormal anchor points.

[0060] In a preferred embodiment, please refer to Figure 4The color edge relationship graph uses color edge anchor points as graph nodes, and establishes graph edges based on same-color contour continuity, page space adjacency, and cross-color common edge relationships. Local offset evidence and color edge topological descriptions are written into the graph nodes, while anchor point coordinate differences, contour tangential correspondences, and color separation participation identifiers are written into the graph edges. The graph neural network performs corresponding message transformations for different graph edge relationships and aggregates messages from adjacent graph nodes in the local coordinate system of the color edge anchor points to obtain node offset states including offset direction, offset continuity state, and evidence credibility. During graph construction, it searches for color edge anchor points belonging to the same color separation and connected sequentially along the contour path along the directed connection records of the pre-press color separation contour, connecting them into same-color contour continuous graph edges. For color edge anchor points with adjacent page coordinates but not belonging to the same contour path, the spatial continuity relationship of their detection areas is used to determine the connection between them. The system establishes adjacency graph edges in the layout space. For anchor points of different color edges that share the same layout boundary but participate in color separation, cross-color common edge graph edges are established according to the direction of the common edge in the local coordinates of the two anchor points. Graph nodes save local offset vectors and evidence credibility for each non-baseline color separation. Graph edges record the rotation mapping, normal orientation relationship, and contour path direction from the local coordinates of the source node to the local coordinates of the target node for each participating color separation. After reading the relationship type, the message transformation selects the corresponding feature combination. Messages with continuous contours of the same color retain the direction consistency information of the same color separation. Layout space adjacency messages retain the displacement change information of the adjacent area. Cross-color common edge messages retain the relative displacement information on both sides of the common boundary. All relationship messages are uniformly converted to the contour tangent and contour normal coordinates of the target graph node before entering the target graph node.

[0061] Furthermore, the graph neural network sets message transformation paths for continuous relationships of the same color contour, adjacency relationships in page space, and cross-color common edge relationships. Based on the contour tangent and contour normal of the color edge anchor points, the offset vectors of adjacent graph nodes are transformed to the local coordinate system of the target graph node. Message gating values ​​are generated based on the credibility of the offset evidence and the correspondence of the contour tangent in the graph edge. The relationship messages output by each message transformation path are aggregated, and the aggregation results are used to update the offset vector, offset continuity state, and credibility of the evidence of the target graph node. In the continuous message transformation path of the same color contour, the offset vectors of the source node and the target node for the same non-reference color are aligned in direction, and the changes along the contour tangent and the changes along the contour normal are separated. The encoding is not performed; in the page space adjacency message transformation path, the page coordinate difference between the source node and the target node is added to the message, so that the processor can distinguish between the smooth displacement changes of adjacent areas and the abrupt changes across areas; in the cross-color common edge message transformation path, the relative order and normal orientation of each color response on both sides of the common edge are added to the message to maintain the color correspondence on the same geometric boundary; the message gating value is jointly generated by the credibility of the source node offset evidence, the credibility of the target node offset evidence, the contour direction difference after local coordinate transformation, and the consistency of the graph edge participation in color separation. The gated relationship messages are summarized according to the relationship type, and then merged with the original state of the target graph node. The updated node offset state continues to serve as the input for the next round of relationship aggregation.

[0062] In this embodiment, the offset vectors of the target graph nodes are aggregated according to the following formula.

[0063] ;

[0064] in, This represents the two-dimensional offset vector of the i-th color edge anchor point after relation aggregation for non-baseline color separation c. This represents the local offset vector before relation aggregation. This represents the local offset vector before the aggregation of relationships between adjacent graph nodes j. This represents a set of relationship types consisting of continuous outlines of the same color, adjacency relationships in page space, and shared edge relationships across colors. This represents the set of adjacent graph nodes connected to the i-th colored edge anchor point through the r-th relation. This represents the message gating value for a path from an adjacent graph node j to a target graph node i, where the corresponding relationship type is r and the non-baseline color separation is c. This represents a two-dimensional transformation matrix that transforms the local coordinates of an adjacent graph node j to the local coordinates of the target graph node i. For example, if the initial lateral offset of the target graph node is 2 image length units and the confidence level of the evidence is 0.8, the lateral offsets of the two adjacent graph nodes after coordinate transformation are 2 and 1 image length units, respectively, corresponding to message gate values ​​of 0.6 and 0.4. After substituting these values, the aggregated lateral offset is approximately 1.78 image length units. The aggregated result maintains the offset direction consistent with the higher confidence evidence and incorporates the continuous change information of adjacent contours.

[0065] Please refer to Figure 5 , Figure 5 The horizontal axis represents the order of the colored anchor points participating in the aggregation of relationships, and the vertical axis represents the horizontal offset length in the local coordinates of the anchor points. The three facets respectively show the offset changes before aggregation, after one aggregation, and after the directed anchor point loop check. Figure 5 correspond , , , , . Figure 5 Before aggregation, the curve retains the offset fluctuations obtained from direct regression of the local detection area. The first aggregation curve tends to be smooth in the section with strong support from adjacent anchor points, indicating that the three types of relationships—continuous same-color outline, adjacent page space, and common cross-color edge—are suppressing isolated fluctuations. After ring verification, the curve further converges to the continuous offset trend, indicating that clockwise and counterclockwise closed propagation reduces the weight of relational edges with directional conflicts, color conflicts, or non-closed cumulative displacement.

[0066] Please refer to Figure 6 , Figure 6 The horizontal axis represents the relationship edges arranged in the recording order in the color edge relationship diagram, and the vertical axis represents the gate value before the relationship message enters the target anchor point. The three types of curves correspond to the same color outline continuous relationship, the page space adjacency relationship, and the cross-color common edge relationship, respectively. Figure 6 Around the aggregation formula in the text Expand It is used to describe the effective proportion of adjacent node j transmitting offset messages to target node i under relation type r and non-baseline color separation c. Its value is affected by the credibility of evidence between source node and target node, the difference in contour direction after local coordinate transformation, and the consistency of participation in color separation. Figure 6 The continuity of the same color outline usually maintains a high and relatively stable gating level, indicating that adjacent anchor points on the same prepress outline path can provide continuous displacement constraints; the fluctuation range of the adjacency relationship in the layout space is relatively large, reflecting that adjacent areas may support stable displacement or be weakened at local abrupt changes; the cross-color common edge relationship is low on some relationship edges, indicating that if there is an inconsistency in normal orientation, conflict in color separation order, or weak color response on both sides of the common boundary, its message will not be forcibly propagated.

[0067] In a preferred embodiment, in the color edge relationship diagram, clockwise and counterclockwise directed anchor point loops are generated from the color edge anchor points on the same closed printing contour. The node offset states of each graph node are sequentially converted to a unified page coordinate system along the two directions and propagated to the starting graph node. The message credibility of the corresponding graph edge is updated according to the difference in the closed state formed by the two propagation paths at the starting graph node. Then, the updated message credibility is used to perform secondary relationship aggregation on each graph node in the directed anchor point loop, and the node offset state obtained from the secondary relationship aggregation is written into the page color registration deviation distribution. In practice, the processor reads the contour direction and anchor point arrangement order from the closed contour record of the pre-press color separation layer, connects adjacent color edge anchor points on the same closed contour sequentially, and specifies a clockwise propagation direction. Then, all graph edge directions are reversed to form a counter-clockwise directed anchor loop. Each propagation transforms the local offset vector of the current graph node to the coordinates of the next graph node through the coordinate transformation matrix of the corresponding graph edge. Then, the propagation state is updated based on the credibility of the offset evidence of the next graph node until the propagation state returns to the starting graph node. The processor compares the color-coded identifier, offset direction, and cumulative displacement change returned by the clockwise path with the corresponding content returned by the counter-clockwise path item by item. The difference in the closure state is returned to the two paths according to the position of the graph edge where the difference occurs. The graph edge that causes direction conflict, color conflict, or non-closed cumulative displacement reduces the message credibility, while the graph edge that maintains the closure retains the original message credibility. The secondary relation aggregation reads the updated graph edge state and recalculates the offset vector of the graph node in the loop.

[0068] In this embodiment, the same colored edge anchor point can belong to multiple directed anchor point loops simultaneously. For example, when a colored edge anchor point is located in the connection area between the outer contour and the inner hollow contour of the graphic, the processor establishes closed propagation paths along the outer contour and the inner contour respectively, and saves an independent closed state record for each propagation path. When multiple directed anchor point loops give the same message credibility adjustment direction for the same graphic edge, the differences in each closed state are merged according to the number of credible nodes in the corresponding loop. When different directed anchor point loops give opposite adjustment directions for the same graphic edge, the bidirectional closed loop residuals and candidate topology correspondence results of the nodes at both ends of the graphic edge are read, and the results are retained with the contour point. The adjustment record matches the closed loop return position and the intersection order; for contours that cannot form a complete directed anchor point loop due to image occlusion, lack of ink, or cropping, no virtual closed edge is generated, but the original continuous image edge of the same color contour is retained and the corresponding contour segment is recorded as an open path. The open path participates in the first relation aggregation but does not participate in the closure state verification; after the second relation aggregation is completed, each image node simultaneously saves the original local offset vector, the first aggregated offset vector, and the offset vector after the loop verification. The color registration deviation distribution of the page uses the offset vector after the loop verification as the basis for continuous offset calculation, and the credibility of evidence at different stages is stored as a judgment record.

[0069] Furthermore, the generation of the color registration deviation distribution includes: projecting the node offset states of each graph node onto a unified page coordinate system according to the color separation identifiers; separating the common translation component, the local residuals that continuously change along the printing outline, and the isolated residuals lacking graph connection support from the node offset states corresponding to the same color separation; forming an overall deviation sub-graph based on the common translation component, forming a regional gradient deviation sub-graph based on the local residuals, and recording the isolated residuals as evidence of discontinuous offset; combining the credibility of the evidence from each graph node to fuse the overall deviation sub-graph and the regional gradient deviation sub-graph, and outputting a color registration deviation judgment result containing offset color separation, spatial range, offset direction, and deviation type. For each non-baseline color separation aggregate participating in color separation, the offset vector in the local coordinates of each graph node is converted into horizontal and vertical components of the unified layout. Based on the credibility of the evidence and the connectivity of the relationship graph, the common translation component covering the current layout is obtained. Then, the common translation component is subtracted from the offset vector of each graph node to obtain the local residual. Local residuals belonging to the same connected subgraph and continuously changing along the printing outline direction are written into the regional gradient deviation subgraph. Local residuals that do not form the same direction support with adjacent graph nodes and exist only in a single detection area are written into the non-continuous offset evidence set. The overall deviation subgraph and the regional gradient deviation subgraph retain their respective color separation identifiers, coordinate ranges, and credibility of evidence.

[0070] Specifically, the overall deviation sub-graph establishes a common displacement record covering the entire printing area based on color separation. When the same non-reference color separation exhibits a consistent directional offset among multiple non-adjacent but topologically reliable color edge anchor points, the processor includes the nodes in the calculation of the common translation component and determines the coverage area of ​​the overall deviation sub-graph based on the printable area of ​​the reachable nodes in the relation graph. The regional gradient deviation sub-graph is established based on the connected components of the relation graph. The processor reads the directional and amplitude changes of the local residuals along the continuous edges of the same-color contour and interpolates the continuously changing node sequence to the corresponding printed contour coverage area. Interpolation only occurs when there are same-color contour connections or adjacent areas in the print space. The process is executed between nodes with connections, without crossing areas of the page that lack image edges; non-continuous offset evidence maintains the original detection area range and does not participate in the spatial extrapolation of the overall deviation sub-image and the regional gradient deviation sub-image; when generating the judgment result, the allowable range of color separation and different page areas corresponding to each color separation is read from the printing task record, and the offset vectors in the overall deviation sub-image and the regional gradient deviation sub-image are compared with the corresponding allowable ranges respectively. The output is the color separation that has shifted, the continuous area that exceeds the range, the horizontal and vertical offset directions of the page, and the overall deviation or regional gradient deviation type. Non-continuous offset evidence is saved as independent verification data along with the judgment result.

[0071] In a preferred embodiment, the deep feature extraction network and the graph neural network are jointly trained. The training data consists of the pre-press color separation layer of the same printing plate, the printing image to be inspected, the correspondence of color edge anchor points, the local offset annotations of each anchor point, and the annotations of plate deviation areas. After the processor constructs a complete color edge relationship graph according to the printing plate, it batch inputs the detection areas into the color response branch and the contour geometry branch. The resulting local offset vectors and the credibility of the offset evidence are used as the initial state of the graph nodes and input into the graph neural network. The training constraints include the correspondence constraints between the color separation response map and the pre-press color separation ink application area, and the local offset vectors and known color separations. The constraints include regression constraints between moving annotations, closed-loop constraints between forward and reverse mappings, offset continuity constraints between adjacent nodes of the same color contour, relative displacement constraints on both sides of the common edge across colors, and distribution constraints between node aggregation results and layout deviation area annotations. All constraints are fed back to the network path that generates the corresponding data. Color appearance change samples mainly update the color separation projection parameters in the color response branch. Known color separation moving samples mainly update the contour geometry branch and offset regression layer. Relationship continuity samples mainly update the graph edge message transformation and message gating parameters. This avoids a single layout-level judgment error from changing all feature paths at the same time.

[0072] In this embodiment, training samples are grouped according to the printing layout version and the original sheet source. Derivative samples formed by color changes, color separation shifts, or image cropping of the same original sheet are kept in the same data group to prevent images with the same contour noise from entering both training and validation data simultaneously. Local offset annotations are generated by the known shift of the color separation response map relative to the reference color separation response map. Actual sheet samples are generated through manual verification of the correspondence between the pre-press color separation contour and the high-confidence detection contour. The layout deviation area annotations are formed by projecting the continuous anchor point offset records of the same color separation. Training data coverage closure. Contours, intersecting contours, adjacent contours, weak color response contours, partially occluded contours, and repeating pattern contours are identified, and each contour retains the states of normal color matching, common translation, and continuous local change. During the verification process, the color edge anchor point positioning results, local offset vectors, the credibility of offset evidence, the state of relation graph nodes, and the final deviation distribution are checked respectively. If the layout-level results do not conform to the annotation, the deviation distribution is traced back to the corresponding graph nodes, graph edges, and detection areas to determine whether the error source belongs to candidate topology assignment, color response generation, local offset regression, or relation message propagation. Then, the corresponding network path is updated using the corresponding type of sample.

[0073] In one online detection embodiment, after receiving the layout version identifier of the current printing task, the processor preloads the prepress color separation layer, color edge anchor point set, color edge topology description, and color edge relationship graph structure. When the printed image to be inspected arrives, only overall coordinate correction, candidate topology correspondence, detection area cropping, local offset inference, and graph node status update are performed. The preloaded graph structure avoids repeatedly extracting the prepress contour for each printed sheet. Each detection area is written into a graph node according to its available state in the printed image to be inspected. Graph nodes that are successfully located and have closed bidirectional mappings enter the complete evidence state, while those that are successfully located but have closed-loop residuals enter the complete evidence state. Larger graph nodes enter the waiting-to-aggregate state, while graph nodes for which no candidate topology is found enter the missing state. The graph neural network performs relational message propagation on the complete evidence state and the waiting-to-aggregate state. The missing state does not send local offset messages but retains the graph edges between it and its neighboring nodes. The color registration deviation distribution of the page is generated only by nodes with actual offset evidence. After processing, the judgment result is bound to the sheet acquisition identifier, and the color edge anchor points, local offset vectors, relational aggregation states, ring check states, and deviation subgraphs involved in the judgment are saved, so that the judgment of the same sheet can be traced back from the page result to the detection area layer by layer.

[0074] In a preferred anomaly handling embodiment, when a candidate topology description forms a similar correspondence with multiple color edge anchors and the joint assignment of adjacent anchors still cannot determine a unique relationship, the processor will form relevant candidates into a local matching subgraph. It will assume each candidate assignment and calculate the consistency of color separation order, the degree of contour closure, and the continuity of page coordinate changes between the candidate and confirmed adjacent graph nodes. Candidate assignments with fewer relationship constraint conflicts will be retained, and their initial evidence credibility will be reduced. When the color edge relationship graph forms multiple disconnected subgraphs due to sheet cutting or large-area occlusion, each connected subgraph will complete relationship aggregation and deviation distribution generation separately. Continuous offsets will not be inferred between different connected subgraphs. Only when multiple connected subgraphs form a common translation component with consistent direction for the same color separation and both have bidirectional closed-loop evidence will they be merged into an overall deviation subgraph. When a detection area has ink spots, scratches, or ink shortages causing local contour breaks, the connection gaps in the candidate topology description will be transmitted to the graph node state. The graph node can receive messages from adjacent nodes, but its local offset evidence will not be sent as a highly credible message, limiting the spread of isolated anomalies along the color edge relationship graph.

[0075] Furthermore, for continuous displacements caused by local stretching or deformation of the printing plate, the processor does not force all graph nodes to be merged into the same common translation component. Instead, it reads the direction of change of local residuals with the position of the printing plate according to the continuous graph edges of the same color contour. When the local residuals of adjacent color edge anchor points change continuously in the unified printing plate coordinate system and the bidirectional closed-loop evidence remains corresponding, the node sequence is retained in the regional gradient deviation subgraph. For color registration deviations caused by changes in the overall position of the printing plate, the same non-reference color separation presents an approximately common offset vector at color edge anchor points distributed in different positions on the printing plate. After the relationship is aggregated, the common translation component covers multiple connected contour areas, and the processor records it as an overall deviation. For color anomalies caused by changes in ink layer concentration, illumination reflection, or printing plate texture, the color appearance characterization changes while the geometric alignment characterization does not form a stable positive and negative contour correspondence. The closed-loop residuals and the credibility of the relationship messages of the relevant nodes cannot support the continuous subgraph, and the processor retains it as non-continuous offset evidence. The three states are distinguished by the same deviation distribution generation process.

[0076] In different printing object embodiments, label and packaging printing uses text outlines, graphic boundaries, and spot color intersection areas in the pre-press color separation file to establish color edge anchor points. Identical patterns in repeating layouts are distinguished by layout coordinates and adjacent anchor point connection paths. Fabric printing uses color separation boundaries of pattern cycle units to establish anchor points. During overall coordinate correction, local fabric deformation is preserved. The adjacent edges in the layout space of the relationship diagram are connected according to their relative positions within the pattern cycle unit, so that the continuous residuals formed by local stretching are included in the regional gradient deviation sub-graph. Digital printing uses the color separation raster layer of the digital original to directly generate color edge topology descriptions and connects the outlines spanning multiple printing areas in the same image content into continuous outline edges of the same color. When the printing object contains spot colors, the spot colors are used as independent color separation identifiers to participate in color response diagram generation and relationship diagram construction. They are not forcibly converted into other color separation combinations. The reference color separation is still determined based on the outline connectivity structure within the current anchor point. The generation logic of local offset vectors, two-way closed-loop confidence, and layout deviation distribution remains consistent.

[0077] In one batch adaptation embodiment, the processor selects detection areas from qualified printed sheets continuously obtained from the current layout that have complete topological correspondence, small bidirectional closed-loop residuals, and consistent relationship graph closure states. It establishes a color appearance sample set for the current batch and uses this set to update the characterization parameters related to brightness, overall color, and ink layer response in the color response branch. Simultaneously, it freezes the geometric relationship parameters in the contour geometry branch, offset regression layer, and graph neural network. Before and after the update, the same set of known color separation movement samples is used to check the local offset vector and deviation distribution. The updated color response parameters are only enabled when the geometric output remains consistent. If a new layout file appears in the current batch, the color edge anchor point set and color edge relationship graph are regenerated, and an independent layout version record is established without using the anchor point identity of the old layout. By restricting color appearance adaptation to the color response path and using color edge topological description, bidirectional closed-loop mapping, and relationship graph continuous state constraints to adapt the samples, the material color and ink layer appearance changes of the current batch can enter the color separation response generation process, while the relative positional relationship defined by the pre-press contour is still used as the basis for color registration deviation judgment.

[0078] In one verification embodiment, when generating the color mismatch determination result, the processor simultaneously establishes a hierarchical reverse index between the deviation region, the layout grid, graph nodes, color edge anchor points, and the detection region. Each deviation region records the continuous graph node set, common translation components, regional gradient residuals, and evidence of discontinuous offsets forming the region. Each layout grid records the graph node identifiers involved in interpolation, the contour path distance from the node to the grid, and the corresponding color separation identifier. When the deviation region is formed by an overall deviation sub-graph, the processor reads graph nodes distributed in different layout positions that support common translation components along the reverse index and checks the bidirectional contour correspondence consistency and local coordinate transformation records of each graph node. When the deviation region is formed by a regional gradient deviation sub-graph, the processor reads graph nodes distributed in different layout positions that support common translation components along the same color index. The continuous outline graph restores the node arrangement order and transforms the local residuals of each node after deducting the common translation component to a unified layout coordinate system. Then, the continuous change trajectory is restored according to the adjacent relationship on the outline path. For the layout grid at the boundary of the deviation area, the interpolation result is retained only when the graph nodes connected to the grid belong to the same relation graph connected component and have the same color identifier. Grids lacking the above connection conditions do not extend the deviation area outward. The verification data also retains the corresponding positions of the candidate topology description and the color edge topology description, the outline point pairs used for forward mapping and reverse mapping, the message gating values ​​of each relation type, and the differences in the closure state of the directed anchor loop, so that any position in the deviation area can be traced back to the specific color edge structure that produced the evidence of local offset.

[0079] Furthermore, when multiple non-reference color separation offset records appear simultaneously on the same printed sheet, the processor establishes node offset status and page registration deviation distribution for each color separation, and maintains the coordinate correspondence of different color separation results through a common reference color separation identifier, without directly merging the local offset vectors of different non-reference color separations; if two non-reference color separations share the same color edge anchor point, the cross-color common edge diagram only transmits the direction correspondence and relative displacement relationship of the common boundary, and each color separation still retains its independent evidence credibility, common translation component, and local residual; if a certain color edge anchor point selects different reference color separations in different local structures, the processor uses pre-press analysis... In the color layer, the connection chain participating in color separation establishes the color separation coordinate transformation path, converts the relative offset of adjacent anchor points to the common benchmark color separation used in the current judgment, and nodes that lack a traceable transformation path do not participate in cross-anchor point color separation merging; after the color separation coordinates are unified, the processor reads the offset direction and spatial range of each color separation in the same page area, writes the results with independent graph connection support into the corresponding color separation judgment record, and retains the records that are only derived from cross-color relationships but lack local contour evidence of this color separation as auxiliary relationship data, thereby maintaining the data closed loop between each color separation judgment and its detection area, contour point correspondence and relationship propagation path.

Claims

1. A method for determining color registration deviation in printing based on deep learning, characterized in that, include: Obtain the pre-press color separation layer and the printing image to be inspected corresponding to the pre-press color separation layer; generate a set of color edge anchor points based on the contour overlap relationship of each pre-press color separation layer; and locate the detection area corresponding to each color edge anchor point in the printing image to be inspected. Local offset evidence of color separation contours is obtained from each detection region through a deep feature extraction network; A color edge relationship graph is constructed based on the continuity relationship of the same color contour and the spatial adjacency relationship between the color edge anchor points. The local offset evidence of adjacent color edge anchor points is aggregated through a graph neural network to obtain the node offset state. Generate the color registration deviation distribution of the page layout based on the offset status of each node and output the color registration deviation judgment result.

2. The printing color registration deviation determination method based on deep learning according to claim 1, characterized in that, The generation of the color edge anchor point set includes: extracting closed contours, intersecting contours and adjacent contours from each prepress color separation layer, projecting the contours of different prepress color separation layers onto a unified page coordinate system, and forming a color edge topology description according to the color separation identifier, contour intersection order, edge direction and adjacency relationship. Select a layout area containing at least two color separation contours and having a definite intersection order from the color edge topology description, and record the center position of the layout area, the color separation identifiers, the edge direction sequence and the contour connection relationship as the color edge anchor point.

3. The printing color registration deviation determination method based on deep learning according to claim 1, characterized in that, The deep feature extraction network includes a color response branch and a contour geometry branch; The color response branch converts the detection area into a color contrast representation and generates each color separation response map based on the color separation participation identifier of the corresponding color edge anchor point; The contour geometry branch extracts gradient direction, edge phase and contour connection structure from the detection area, and performs cross-attention fusion on each color response map according to the correspondence between the participating color divisions; The fused features are input into the offset regression layer to obtain the local offset vectors of each non-reference color segment relative to the reference color segment and the credibility of the offset evidence.

4. The printing color registration deviation determination method based on deep learning according to claim 2, characterized in that, The color edge relationship diagram uses the color edge anchor point as the graph node, and establishes graph edges based on the same color outline continuity relationship, the page space adjacency relationship, and the cross-color common edge relationship respectively; Write the local offset evidence and the color edge topology description into the graph node, and write the anchor point coordinate difference, the contour tangential correspondence and the color separation identifier into the graph edge; The graph neural network performs corresponding message transformations for different graph edge relationships and aggregates the messages of adjacent graph nodes in the local coordinate system of the colored edge anchor point to obtain the node offset state, which includes the offset direction, offset continuity state, and evidence credibility.

5. The printing color registration deviation determination method based on deep learning according to claim 1, characterized in that, Locating the detection area corresponding to each color edge anchor point includes: extracting candidate edges in the printed image to be inspected, and generating candidate topological descriptions according to the direction order, color response order, and intersection connection relationship of the candidate edges; The candidate topology description is matched with the color edge topology description of the color edge anchor point, and a local coordinate system of anchor point composed of contour tangent and contour normal is established for the candidate edges that form the correspondence. Based on the local coordinate system mapping detection area of ​​the anchor point, when there are multiple candidate correspondences, joint assignment is completed by combining the arrangement order of adjacent colored edge anchor points and the contour connection relationship.

6. The printing color registration deviation determination method based on deep learning according to claim 3, characterized in that, The training of the color response branch and the contour geometry branch includes: constructing a first sample pair with the same geometric position relationship but different color brightness or ink layer response, and a second sample pair with similar color appearance but relatively shifted color separation contours; The color appearance representation and geometric alignment representation in the sample are encoded separately, and the color appearance representation is exchanged between the first sample pairs and recombined with the original geometric alignment representation; The local offset vectors before and after the constraint exchange remain consistent, and the geometric alignment representation is constrained by the color-separation movement annotation of the second sample pair, so that the local offset evidence in the detection stage is jointly generated by the geometric alignment representation and the color edge topology description.

7. The printing color registration deviation determination method based on deep learning according to claim 4, characterized in that, The generation of the local offset evidence further includes: inputting the reference image region of the color edge anchor point and the corresponding detection region into the contour geometric branch with shared parameters in a forward order and a reverse order, respectively, to obtain a forward offset vector from the reference image region to the detection region and a reverse offset vector from the detection region to the reference image region. The reference contour points are mapped sequentially using the forward offset vector and the reverse offset vector. The credibility of the offset evidence is generated based on the closed-loop residual between the mapped contour points and the original reference contour points and the bidirectional contour correspondence consistency. The credibility of the offset evidence and the forward offset vector are written together into the corresponding graph node.

8. The printing color registration deviation determination method based on deep learning according to claim 7, characterized in that, The graph neural network sets message transformation paths for the same-color contour continuity relationship, the page space adjacency relationship, and the cross-color common edge relationship, respectively; Based on the contour tangent and contour normal of the color edge anchor point, the offset vectors of adjacent graph nodes are transformed to the local coordinate system of the target graph node; Based on the credibility of the offset evidence and the contour tangential correspondence in the graph edge, a message gating value is generated. The relational messages output by each message transformation path are aggregated, and the offset vector, offset continuity state, and credibility of the evidence of the target graph node are updated using the aggregation result.

9. The printing color registration deviation determination method based on deep learning according to claim 5, characterized in that, In the color edge relationship diagram, clockwise directed anchor point loops and counterclockwise directed anchor point loops are generated from the color edge anchor points on the same closed printing contour. The node offset state of each diagram node is sequentially converted to the unified layout coordinate system along the two directions and propagated to the starting diagram node. The message credibility of the corresponding graph edge is updated based on the difference in the closed state formed by the two propagation paths at the starting graph node. Then, the updated message credibility is used to perform secondary relationship aggregation on each graph node in the directed anchor loop, and the node offset state obtained by the secondary relationship aggregation is written into the page color deviation distribution.

10. The printing color registration deviation determination method based on deep learning according to claim 9, characterized in that, The generation of the color matching deviation distribution of the layout includes: projecting the node offset state of each image node onto the unified layout coordinate system according to the color matching identifier; Separate common translation components, local residuals that vary continuously along the printing outline, and isolated residuals that lack graph connectivity from the node offset states corresponding to the same color separation. An overall deviation submap is formed based on the common translation components, a regional gradual deviation submap is formed based on the local residuals, and the isolated residuals are recorded as evidence of discontinuous offset. By combining the credibility of evidence from each graph node with the overall deviation subgraph and the regional gradient deviation subgraph, a color deviation determination result is output, which includes offset color separation, spatial range, offset direction, and deviation type.