A three-dimensional cartoon character reconstruction method based on deep learning

Abstract

This invention discloses a deep learning-based method for reconstructing 3D anime characters, comprising the following steps: acquiring a single anime character image and generating a normalized character image; extracting the contour centerline, bandwidth, direction, curvature, opening / closing state, and attachment state to form a contour topological phase band set; constructing a contour topology map containing extension, attachment, wrapping, and occlusion relationships; generating an initial 3D character geometric model according to the attachment phase, outward expansion phase, and wrapping phase; generating topological phase band deviation results through projection comparison; and performing local surface rewriting based on the deviation results to obtain the 3D anime character geometric model. This invention employs deep learning and contour topological phase band modeling to achieve 3D anime character reconstruction, possessing advantages such as high contour fidelity, accurate structural hierarchy, and strong geometric stability.

Description

Technical Field

[0001] This invention relates to the field of 3D modeling technology, and in particular to a method for reconstructing 3D animation characters based on deep learning. Background Technology

[0002] With the development of computer vision, computer graphics, and deep learning technologies, the technique of generating 3D character models from single images has been widely applied in scenarios such as animation production, virtual display, digital entertainment, and interactive modeling. Existing technologies typically input 2D character images into convolutional neural networks, encoder-decoder networks, or implicit 3D generative networks to extract appearance, contour, or semantic features from the images, and then output the corresponding 3D character model. Some techniques also combine foreground segmentation, contour extraction, projection comparison, vertex adjustment, or local surface optimization to improve the structural integrity and detail of the 3D model, thereby achieving automatic conversion from 2D animation images to 3D character models.

[0003] However, most existing technologies focus on recovering the general three-dimensional shape of a character from a two-dimensional image, typically using contour information as a general edge feature or a basis for subsequent corrections. They lack a structured expression of the extension, attachment, wrapping, and occlusion relationships within the contours of anime characters, making it difficult to accurately reflect the hierarchical organization and local geometric semantics of the character's contour. Furthermore, existing technologies often employ a uniform geometric generation method for different contour regions, lacking differentiated generation mechanisms based on attachment phase, outward expansion phase, and wrapping phase. They also lack closed-loop comparison and local surface rewriting mechanisms based on center, bandwidth, direction, curvature, and state, easily leading to contour distortion, hierarchical misalignment, local structural collapse, and insufficient fidelity in the anime style of the generated 3D anime character models.

[0004] Therefore, how to provide a method for reconstructing 3D animation characters based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for reconstructing 3D animation characters based on deep learning. This invention uses deep learning and contour topology phase band modeling to achieve 3D animation character reconstruction, which has the advantages of high contour fidelity, accurate structural hierarchy and strong geometric stability.

[0006] A method for reconstructing 3D animation characters based on deep learning according to an embodiment of the present invention includes the following steps: The process involves acquiring a single anime character image, performing foreground separation, image normalization, line art enhancement, and contour distance transformation to generate a normalized character image. The normalized character image is input into the contour topology phase band extraction network to extract the contour centerline, bandwidth, direction, curvature, opening and closing state and attachment state, and generate a contour topology phase band set. A contour topology map containing continuation, attachment, wrapping, and occlusion relationships is established based on the contour topology phase band set. The normalized character image and contour topology map are input into the phase-separated geometry generation network to generate an initial 3D character geometric model according to the attached phase, the outward phase, and the wrapping phase. Perform projection mapping on the initial 3D character geometry model, compare it with the contour topology phase band set in terms of band center, bandwidth, direction, curvature and state correspondence, and generate topology phase band deviation results; Based on the topological phase band deviation results, a local surface rewrite is performed on the initial 3D character geometric model to obtain the 3D animation character geometric model.

[0007] Optionally, the generation of the normalized character image specifically includes: The process involves acquiring a single anime character image, performing pixel validity checks, alpha channel parsing, and color space unification to generate the original character image. Perform foreground separation on the original character image, output a foreground probability map, perform consistency fusion with the transparent region marker map or the background candidate region map, calculate the fused foreground response value, and obtain the foreground mask map; Based on the foreground mask, the minimum bounding region of the foreground is extracted. The original character image is then cropped, centered, translated, and scaled to obtain a cropped character image. Boundary padding is then performed based on the aspect ratio of the minimum bounding region of the foreground and the preset target side length to generate a scaled character image. Perform image normalization on scale-uniform character images to generate normalized image results; The image normalization result is input into the line art enhancement network to extract the character's outer contour response, local boundary response, and high curvature inflection point response. The background region response is suppressed based on the foreground mask to generate the line art enhancement image. Based on the enhanced line art image, closed and open contour regions are extracted. Contour distance transformation is performed to obtain inner and outer contour distance maps, respectively. These maps are then encapsulated with the image normalization result, foreground mask map, and enhanced line art image in channel order to generate a normalized character image.

[0008] Optionally, the generation of the contour topology phase band set specifically includes: Normalized character images are input into a contour topology phase band extraction network in channel order. Joint encoding is performed on the normalized character images to extract multi-scale contour response features, region boundary features, and distance distribution features, generating a contour extraction input feature map. Perform centerline extraction processing on the contour extraction input feature map to generate contour centerline results; Based on the contour centerline results and the contour extraction input feature map, bandwidth extraction processing is performed to generate bandwidth results that correspond one-to-one with the contour centerline. Based on the contour centerline results, bandwidth results, inner contour distance map and outer contour distance map, perform direction and curvature extraction processing to generate contour geometric description results; Based on the contour geometry description results and the foreground mask, state determination processing is performed to obtain the opening / closing state and attachment state corresponding to each contour centerline, and generate contour state results. The contour centerline results, bandwidth results, contour geometric description results, and contour state results are associated and encapsulated by contour segments. Adjacent contour segments with the same attributes are merged, and contour segments with conflicting attributes are segmented by boundary, generating a contour topology phase band set.

[0009] Optionally, the centerline extraction process includes refining and clustering the continuous contour response region and suppressing breakpoints.

[0010] Optionally, the generation of the contour topology map specifically includes: Read the contour topology phase band set, establish a contour segment index according to the contour center line, bandwidth result, contour geometric description result and contour state result corresponding to each contour topology phase band, and generate a contour segment candidate association set based on the positional continuity of the contour center line in the normalized character image and the segment affiliation relationship. Based on the contour segment index and the candidate association set of contour segments, a continuation relationship determination is performed to obtain the continuation relationship set; Based on the contour fragment index, foreground mask map, and contour state results, an attachment relationship determination is performed to obtain an attachment relationship set. Based on the contour segment index, inner contour distance map, outer contour distance map and contour geometric description results, the wrapping relationship determination is performed to obtain the wrapping relationship set; Occlusion relationship determination is performed based on the contour fragment index, line art enhancement image, foreground mask image and contour state results to obtain an occlusion relationship set; Write the sets of extension relations, attachment relations, wrapping relations, and occlusion relations into the relation index table according to the relation type. Perform relation conflict adjudication on the contour fragments that simultaneously satisfy multiple relations to obtain the relation adjudication result set. Using the contour topology phase band set as the node set, and the continuation relationship, attachment relationship, wrapping relationship and occlusion relationship in the relationship adjudication result set as the edge set, the contour topology graph is generated by associating and encapsulating the edges according to edge type, edge start and end segments and corresponding position indexes.

[0011] Optionally, the generation of the initial 3D character geometric model specifically includes: Based on the phase-splitting geometric generation network, the network extracts the overall appearance features, local texture features, and spatial distribution features of the character from the normalized character image. It also extracts the node attribute features, edge relationship features, and relationship transfer features from the contour topology graph. The network is then aligned and fused according to the position index relationship in the contour topology phase band set to generate a geometric generation input feature set. Based on the geometrically generated input feature set, phase determination is performed on each contour topology phase band in the contour topology map to generate a phase annotation result set; Based on the phase annotation result set and the geometric generation input feature set, a phase sub-generation unit is established to obtain the attached phase local surface result, the expanded phase local surface result, and the wrapped phase local surface result. The attached phase local surface results, the extended phase local surface results, and the wrapped phase local surface results are processed according to the set of continuation relationships, attachment relationships, wrapping relationships, and occlusion relationships in the contour topology map. Adjacency stitching, boundary smoothing, and hierarchical adjudication are performed to generate the phase-separated surface stitching results. Based on the phase-separated surface splicing results, the basic character mesh skeleton is established. Continuous splicing is performed on local surfaces with a continuation relationship, connection and fixation are performed on local surfaces with an attachment relationship, enclosing and nesting are performed on local surfaces with a wrapping relationship, and front-to-back hierarchical sorting and occlusion preservation are performed on local surfaces with an occlusion relationship to generate the initial character mesh result. Based on the initial character mesh results and the image normalization results in the normalized character image, region uniformity processing is performed. Mesh subdivision density allocation, boundary orientation straightening, and local curvature buffering are performed on the character outer contour region, local boundary region, and auxiliary structure region, respectively, to generate geometrically uniform results. The geometrically consistent results are associated and encapsulated according to vertex position, surface segment, segment affiliation, and phase source to obtain the initial 3D character geometric model.

[0012] Optionally, the phase generation unit performs attachment generation processing, expansion generation processing, and enclosure generation processing on the attached phase, the outward expansion phase, and the enclosed phase, respectively.

[0013] Optionally, the generation of the topological phase band deviation result specifically includes: Read the initial 3D character geometry model, establish a projection input set based on vertex position, surface segment, segment affiliation relationship and phase source, determine the viewpoint alignment parameters by combining the position index relationship in the normalized character image, and generate a projection input alignment set; Perform projection mapping on the projection input alignment set, and generate the projection contour region, projection center trajectory and projection bandwidth distribution based on the surface fragment boundaries, phase source and front and back hierarchy order in the initial 3D character geometry model, to obtain the projection topology phase band set; The projected topological phase band set is registered with the contour topological phase band set to generate a set of topological phase band alignment results. Based on the topological phase band alignment result set, band center comparison is performed, and the center trajectory position deviation, extension length deviation, and endpoint connectivity deviation of each corresponding segment are calculated segment by segment to obtain the band center deviation result; Based on the topological phase band alignment result set, bandwidth comparison, direction comparison and curvature comparison are performed. The differences in the inner and outer width distribution, local orientation difference and bending change difference of each corresponding segment are calculated segment by segment to obtain the bandwidth deviation result, direction deviation result and curvature deviation result. Based on the set of topological phase band alignment results, a state comparison is performed to determine the consistency of the opening and closing states, the consistency of the attachment states, and the consistency of the continuation, attachment, wrapping, and occlusion relationships inherited from the contour topology graph of each corresponding segment, and to obtain the state deviation results. The center deviation results, bandwidth deviation results, direction deviation results, curvature deviation results, and state deviation results are associated and summarized according to the contour segments. Deviation propagation and integration are performed on continuous segments, and relationship priority adjudication is performed on conflicting segments to generate a segment-level deviation set. The fragment-level deviation set is backfilled and encapsulated according to the contour centerline, bandwidth result, contour geometric description result and contour state result in the contour topology phase band set to obtain the topology phase band deviation result.

[0014] Optionally, the registration is performed according to the position index relationship, the segment belonging relationship and the phase source. Neighborhood search is performed to supplement the registration for projection segments that have not formed a corresponding relationship, and minimum position deviation is selected for projection segments with multiple candidate correspondences.

[0015] Optionally, the generation of the geometric model of the 3D anime character specifically includes: Read the initial 3D character geometric model and topological phase band deviation results, establish a local surface write-back input set according to the position index relationship of contour segments, segment belonging relationship and phase source, and collect the deviations of each surface segment based on the center deviation result, bandwidth deviation result, direction deviation result, curvature deviation result and state deviation result to generate a local surface write-back candidate set; Based on the local surface writeback candidate set, perform writeback trigger determination and generate a writeback trigger result set; Establish a sub-action write-back result set based on the write-back trigger result set and the phase source; Write the set of sub-action writeback results back to the corresponding surface fragment in the initial 3D character geometric model to generate local surface update results; Based on the local surface update results, a consistency check is performed after writing back. The updated surface segment is regenerated with the projection contour region, projection center trajectory and projection bandwidth distribution, and compared with the contour topology phase band set to obtain the set of deviation results after writing back. Based on the set of deviation results after writeback, perform writeback termination determination and generate a set of writeback termination results; Based on the write-back termination result set, all surface fragments that have completed write-back retention, achieved the target after another write-back, and stabilized and frozen are uniformly encapsulated, and the output includes updated vertex positions, updated surface fragments, updated fragment affiliation relationships, and updated phase sources.

[0016] The beneficial effects of this invention are: This invention constructs a complete technical chain consisting of a standardized character image, a contour topology phase band set, a contour topology map, phase-separated geometry generation, topology phase band deviation results, and local surface rewriting. This transforms the contour structure information in a single anime character image from merely a common edge feature in 3D reconstruction into a core constraint driving 3D geometry generation and correction throughout the entire modeling process. Compared to existing technologies, this invention more accurately maintains consistency in contour center, bandwidth distribution, direction changes, curvature changes, and open / closed and attached states within the character's outer contour, local boundaries, and attached structural regions. This significantly improves the performance of 3D anime character geometric models in terms of contour integrity, hierarchical accuracy, and local structural stability.

[0017] Furthermore, this invention establishes continuation, attachment, wrapping, and occlusion relationships on the contour topology phase band set, and performs differentiated geometric generation processing on the attached phase, outward expansion phase, and wrapping phase in the phase-separated geometry generation network. Then, it combines the topology phase band deviation results to perform local surface rewriting on the initial 3D character geometric model. This can effectively reduce the contour distortion, hierarchical misalignment, local structural collapse, and weakening of animation style that are common in the prior art. The generated 3D animation character geometric model can maintain the overall style characteristics of the original animation character image while having higher geometric accuracy, contour fidelity, and structural coordination. Therefore, it is more suitable for 3D modeling of animation characters, virtual display, digital content production, and subsequent character-driven applications. Attached Figure Description

[0018] 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 1This is a flowchart of a deep learning-based 3D animation character reconstruction method proposed in this invention; Figure 2 This is a flowchart illustrating the generation of contour topology phase band sets and the construction of contour topology graphs in a deep learning-based 3D animation character reconstruction method proposed in this invention. Figure 3 This is a flowchart illustrating the process of generating a geometric model of a 3D anime character through local surface rewriting, a method for reconstructing 3D anime characters based on deep learning, as proposed in this invention. Detailed Implementation

[0019] 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.

[0020] refer to Figures 1-3 A method for reconstructing 3D animation characters based on deep learning includes the following steps: The process involves acquiring a single anime character image, performing foreground separation, image normalization, line art enhancement, and contour distance transformation to generate a normalized character image. The normalized character image is input into the contour topology phase band extraction network to extract the contour centerline, bandwidth, direction, curvature, opening and closing state and attachment state, and generate a contour topology phase band set. A contour topology map containing continuation, attachment, wrapping, and occlusion relationships is established based on the contour topology phase band set. The normalized character image and contour topology map are input into the phase-separated geometry generation network to generate an initial 3D character geometric model according to the attached phase, the outward phase, and the wrapping phase. Perform projection mapping on the initial 3D character geometry model, compare it with the contour topology phase band set in terms of band center, bandwidth, direction, curvature and state correspondence, and generate topology phase band deviation results; Based on the topological phase band deviation results, a local surface rewrite is performed on the initial 3D character geometric model to obtain the 3D animation character geometric model.

[0021] In this embodiment, the generation of standardized character images specifically includes: The process involves acquiring a single anime character image, performing pixel validity checks, alpha channel parsing, and color space unification to generate the original character image. The color space unification processing includes converting a single anime character image with a transparent channel into a three-channel color image and a transparent region marker map, and converting a single anime character image without a transparent channel into a three-channel color image and a background candidate region map; Perform foreground separation on the original character image, output a foreground probability map, perform consistency fusion with the transparent region marker map or the background candidate region map, calculate the fused foreground response value, and obtain the foreground mask map; Foreground separation involves inputting the original character image into a foreground separation network consisting of an encoder, a feature aggregation layer, and a decoding output layer. The encoder extracts multi-scale image features from the main character region, edge transition region, and background region. The feature aggregation layer upsamples, aligns, and fuses the image features at different scales to form a fused feature map that combines overall semantic information with local boundary information. The decoding output layer performs pixel-wise response mapping on the fused feature map, outputting the foreground response value corresponding to each pixel position. The value is then normalized to a preset probability range to obtain a foreground probability map that represents the confidence level of each pixel belonging to the character's foreground. The foreground response value is calculated by weighting and summing the foreground response value output by the foreground separation network and the auxiliary response value corresponding to the transparent region marker map or the background candidate region map according to the preset fusion weights. Based on the foreground mask, the minimum bounding region of the foreground is extracted. The original character image is then cropped, centered, translated, and scaled to obtain a cropped character image. Boundary padding is then performed based on the aspect ratio of the minimum bounding region of the foreground and the preset target side length to generate a scaled character image. Perform image normalization on scale-uniform character images to generate normalized image results; Image normalization includes intensity normalization of color channels, position normalization of spatial coordinates, and background uniform filling of pixels outside the foreground region. The color normalization value is calculated by subtracting the minimum pixel value of the current color channel from the original pixel value of the current color channel, and then dividing by the difference between the maximum pixel value and the minimum pixel value of the current color channel. The image normalization result is input into the line art enhancement network to extract the character's outer contour response, local boundary response, and high curvature inflection point response. The background region response is suppressed based on the foreground mask to generate the line art enhancement image. The line art enhancement response is calculated by multiplying the foreground mask value by the weighted sum of the character's outer contour response, local boundary response, and high curvature inflection point response; Based on the enhanced line art image, closed contour regions and open contour regions are extracted. Contour distance transformation is performed to obtain inner contour distance map and outer contour distance map respectively. These are then encapsulated with the image normalization result, foreground mask map and enhanced line art image in channel order to generate a normalized character image. The contour distance value is calculated based on the Euclidean distance between the current pixel and the nearest contour pixel in the enhanced line drawing image.

[0022] In this embodiment, the generation of the contour topology phase band set specifically includes: Normalized character images are input into a contour topology phase band extraction network in channel order. Joint encoding is performed on the normalized character images to extract multi-scale contour response features, region boundary features, and distance distribution features, generating a contour extraction input feature map. Perform centerline extraction processing on the contour extraction input feature map to generate contour centerline results; Based on the contour centerline results and the contour extraction input feature map, bandwidth extraction processing is performed to generate bandwidth results that correspond one-to-one with the contour centerline. The bandwidth extraction process calculates the inner and outer distances from both sides of the contour centerline to the adjacent boundaries, and performs smoothing constraints and anomaly truncation on local abrupt regions. Based on the contour centerline results, bandwidth results, inner contour distance map and outer contour distance map, perform direction and curvature extraction processing to generate contour geometric description results; The direction and curvature extraction process calculates the local tangent change trend and bending change trend along the extension path of the contour centerline to obtain the direction and curvature results corresponding to each contour centerline. The results are then aligned with the bandwidth results to obtain the contour geometric description results. Based on the contour geometry description results and the foreground mask, state determination processing is performed to obtain the opening / closing state and attachment state corresponding to each contour centerline, and generate contour state results. The state determination includes determining the connectivity between the two ends of the contour, the adjacent coverage of the contour, and the connection between the contour and the main area. The contour centerline results, bandwidth results, contour geometric description results, and contour state results are associated and encapsulated by contour segments. Adjacent contour segments with the same attributes are merged, and contour segments with conflicting attributes are segmented by boundary, generating a contour topology phase band set.

[0023] In this embodiment, the centerline extraction process includes refining and clustering the continuous contour response region and suppressing breakpoints. The process involves performing layer-by-layer convolutional refinement on high-response contour regions to obtain a central response map representing the orientation of the main contour. Based on the continuous distribution of contours in the enhanced line drawing image, the wide contour response regions in the central response map are contracted towards the center along the local main direction, so that the multi-pixel wide responses within the same contour band gradually aggregate into single-pixel wide or narrow-band central trajectories, forming a refined aggregation result. Combining the foreground mask map, the inner contour distance map, and the outer contour distance map, bridging and supplementation are performed on broken positions where the interval between adjacent contour segments is less than the preset connectivity threshold and the direction changes continuously. Suppression and deletion are performed on isolated short segments with a length lower than the preset retention threshold, a sudden change in direction exceeding the preset limit, and no connectivity support with the main body region to eliminate noisy breakpoints and false center lines. The retained continuous center trajectories are classified into the outer contour region of the character, the local boundary region, and the auxiliary structure region, respectively, to obtain the outer contour center line of the character, the local boundary center line, and the auxiliary structure center line, which are then integrated into the contour center line result.

[0024] In this embodiment, the generation of the contour topology map specifically includes: Read the contour topology phase band set, establish a contour segment index according to the contour center line, bandwidth result, contour geometric description result and contour state result corresponding to each contour topology phase band, and generate a contour segment candidate association set based on the positional continuity of the contour center line in the normalized character image and the segment affiliation relationship. Based on the contour segment index and the candidate association set of contour segments, a continuation relationship determination is performed to obtain the continuation relationship set; Among them, the continuity relationship determination establishes a continuity relationship for adjacent contour segments that are adjacent in the first and last positions, have continuous directional results, continuous curvature changes, and consistent opening and closing states. Based on the contour fragment index, foreground mask map, and contour state results, an attachment relationship determination is performed to obtain an attachment relationship set. The attachment relationship determination establishes an attachment relationship between a contour segment that is connected to the boundary of the main area at one or both ends, is in the attachment state, and whose bandwidth result maintains a continuous transition within the connection area and the connected contour segment. Based on the contour segment index, inner contour distance map, outer contour distance map and contour geometric description results, the wrapping relationship determination is performed to obtain the wrapping relationship set; Among them, the enclosing relationship determination establishes an enclosing relationship for contour segments that form an outer encirclement in spatial location, have an enclosing trend in directional results, have continuously changing curvature results on the enclosing side, and have inner and outer distance distributions that satisfy the enclosing level. Occlusion relationship determination is performed based on the contour fragment index, line art enhancement image, foreground mask image and contour state results to obtain an occlusion relationship set; The occlusion relationship determination establishes an occlusion relationship for contour segments that overlap at the projection position, whose contour continuity is interrupted by another contour segment, whose foreground response intensity has a difference between the front and back layers, and whose open end points to the occluded area. Write the sets of extension relations, attachment relations, wrapping relations, and occlusion relations into the relation index table according to the relation type. Perform relation conflict adjudication on the contour fragments that simultaneously satisfy multiple relations to obtain the relation adjudication result set. Among them, the adjudication of relationship conflicts shall be carried out in the following order: extension relationship first, attachment relationship second, wrapping relationship third, and obstruction relationship last. Using the contour topology phase band set as the node set, and the continuation relationship, attachment relationship, wrapping relationship and occlusion relationship in the relationship adjudication result set as the edge set, the contour topology graph is generated by associating and encapsulating the edge type, edge start and end segments and corresponding position indexes. In the contour topology map, each edge is generated by the corresponding judgment results among the spatial adjacency results, direction continuity results, curvature continuity results, connectivity support results, distance distribution results, and occlusion truncation results between contour segments, maintaining the original position index relationship with the normalized character image.

[0025] In this embodiment, the generation of the initial 3D character geometric model specifically includes: Based on the phase-splitting geometric generation network, the network extracts the overall appearance features, local texture features, and spatial distribution features of the character from the normalized character image. It also extracts the node attribute features, edge relationship features, and relationship transfer features from the contour topology graph. The network is then aligned and fused according to the position index relationship in the contour topology phase band set to generate a geometric generation input feature set. Based on the geometrically generated input feature set, phase determination is performed on each contour topology phase band in the contour topology map to generate a phase annotation result set; The phase determination is based on the contour state results, attachment relationship set, wrapping relationship set, occlusion relationship set, and the direction and curvature results in the contour geometric description results, respectively determining the attachment phase, outward expansion phase, and wrapping phase corresponding to each contour topological phase band. Based on the phase annotation result set and the geometric generation input feature set, a phase sub-generation unit is established to obtain the attached phase local surface result, the expanded phase local surface result, and the wrapped phase local surface result. The attached phase local surface results, the extended phase local surface results, and the wrapped phase local surface results are processed according to the set of continuation relationships, attachment relationships, wrapping relationships, and occlusion relationships in the contour topology map. Adjacency stitching, boundary smoothing, and hierarchical adjudication are performed to generate the phase-separated surface stitching results. Based on the phase-separated surface splicing results, the basic character mesh skeleton is established. Continuous splicing is performed on local surfaces with a continuation relationship, connection and fixation are performed on local surfaces with an attachment relationship, enclosing and nesting are performed on local surfaces with a wrapping relationship, and front-to-back hierarchical sorting and occlusion preservation are performed on local surfaces with an occlusion relationship to generate the initial character mesh result. Based on the initial character mesh results and the image normalization results in the normalized character image, region uniformity processing is performed. Mesh subdivision density allocation, boundary orientation straightening, and local curvature buffering are performed on the character outer contour region, local boundary region, and auxiliary structure region, respectively, to generate geometrically uniform results. The geometrically consistent results are associated and encapsulated according to vertex position, surface segment, segment affiliation, and phase source to obtain the initial 3D character geometric model.

[0026] In this embodiment, the phase generation unit performs attachment generation processing, expansion generation processing, and encirclement generation processing on the attached phase, the outward expansion phase, and the enclosed phase, respectively. Specifically, the process involves performing an attachment generation process on the contour topology phase band marked as the attached phase, which generates the attached phase local surface result by constraining the corresponding geometric region to continuously extend along the main surface and suppressing abrupt changes in the normal direction. The process also involves performing an outward expansion generation process on the contour topology phase band marked as the outward expansion phase, which generates the outward expansion phase local surface result by constraining the corresponding geometric region to jointly expand along the contour's outer normal and tangential directions while maintaining boundary continuity. Finally, the process involves performing a wrapping generation process on the contour topology phase band marked as the wrapping phase, which generates the wrapping phase local surface result by constraining the corresponding geometric region to form a bending and closing trend around the wrapped region while maintaining continuity of the inner and outer layers.

[0027] In this embodiment, the generation of the topological phase band deviation result specifically includes: Read the initial 3D character geometry model, establish a projection input set based on vertex position, surface segment, segment affiliation relationship and phase source, determine the viewpoint alignment parameters by combining the position index relationship in the normalized character image, and generate a projection input alignment set; Perform projection mapping on the projection input alignment set, and generate the projection contour region, projection center trajectory and projection bandwidth distribution based on the surface fragment boundaries, phase source and front and back hierarchy order in the initial 3D character geometry model, to obtain the projection topology phase band set; The projected topological phase band set is registered with the contour topological phase band set to generate a set of topological phase band alignment results. Based on the topological phase band alignment result set, band center comparison is performed, and the center trajectory position deviation, extension length deviation, and endpoint connectivity deviation of each corresponding segment are calculated segment by segment to obtain the band center deviation result; Based on the topological phase band alignment result set, bandwidth comparison, direction comparison and curvature comparison are performed. The differences in the inner and outer width distribution, local orientation difference and bending change difference of each corresponding segment are calculated segment by segment to obtain the bandwidth deviation result, direction deviation result and curvature deviation result. Based on the set of topological phase band alignment results, a state comparison is performed to determine the consistency of the opening and closing states, the consistency of the attachment states, and the consistency of the continuation, attachment, wrapping, and occlusion relationships inherited from the contour topology graph of each corresponding segment, and to obtain the state deviation results. The center deviation results, bandwidth deviation results, direction deviation results, curvature deviation results, and state deviation results are associated and summarized according to the contour segments. Deviation propagation and integration are performed on continuous segments, and relationship priority adjudication is performed on conflicting segments to generate a segment-level deviation set. The fragment-level deviation set is backfilled and encapsulated according to the contour centerline, bandwidth result, contour geometric description result and contour state result in the contour topology phase band set to obtain the topology phase band deviation result. The topological phase band deviation results include band center deviation results, bandwidth deviation results, direction deviation results, curvature deviation results, state deviation results, as well as the position index relationship and phase source of the corresponding contour segments.

[0028] In this embodiment, registration is performed according to position index relationship, segment belonging relationship and phase source. Neighborhood search is performed to supplement registration for projection segments that have not formed a corresponding relationship. Minimum position deviation is selected for projection segments with multiple candidate correspondences. The neighborhood search registration is based on the projection center trajectory of the projection segment. It gradually expands the search of candidate segments in the contour topology phase band set within the surrounding preset search range. It prioritizes screening contour segments that have the same segment affiliation and the same phase source as the projection segment. It compares the proximity of the center trajectory, the continuity of the bandwidth distribution, the consistency of the direction, the similarity of the curvature change, and the degree of endpoint connectivity matching. It supplements the candidate segments that meet the preset continuity conditions to establish corresponding relationships. The minimum position deviation is preferably calculated by calculating the center trajectory position deviation, endpoint position deviation and coverage area deviation between each candidate segment and the projection segment. The position deviations are sorted and decided according to a preset priority order. The candidate segment with the smallest center trajectory position deviation, the smallest endpoint position deviation and the highest coverage area overlap is retained as the final corresponding segment, and the corresponding relationship of the remaining candidate segments is removed.

[0029] In this embodiment, the generation of the geometric model of the 3D animation character specifically includes: Read the initial 3D character geometric model and topological phase band deviation results, establish a local surface write-back input set according to the position index relationship of contour segments, segment belonging relationship and phase source, and collect the deviations of each surface segment based on the center deviation result, bandwidth deviation result, direction deviation result, curvature deviation result and state deviation result to generate a local surface write-back candidate set; Based on the local surface writeback candidate set, perform writeback trigger determination and generate a writeback trigger result set; The write-back trigger determination includes marking surface segments whose center deviation results exceed the center deviation threshold, bandwidth deviation results exceed the width deviation threshold, direction deviation results exceed the direction deviation threshold, curvature deviation results exceed the bending deviation threshold, and state deviation results contain inconsistent opening and closing states or inconsistent attachment states as segments to be written back. Establish a sub-action write-back result set based on the write-back trigger result set and the phase source; The set of write-back results includes: performing attachment write-back on segments to be written back with the phase source being the attached phase; performing outward expansion write-back on segments to be written back with the phase source being the outward expansion phase; performing enclosing write-back on segments to be written back with the phase source being the wrapping phase; performing center trajectory correction on segments corresponding to the center deviation results; performing boundary width adjustment on segments corresponding to the bandwidth deviation results; performing direction rotation correction on segments corresponding to the direction deviation results; performing bending buffer correction on segments corresponding to the curvature deviation results; and performing connectivity correction on segments corresponding to the state deviation results. Write the set of sub-action writeback results back to the corresponding surface fragment in the initial 3D character geometric model to generate local surface update results; The write-back includes performing boundary continuum write-back on adjacent surface segments with a continuation relationship, performing connection transition write-back on adjacent surface segments with an attachment relationship, performing inner and outer layer write-back on adjacent surface segments with a wrapping relationship, and performing front and back retention write-back on adjacent surface segments with an occlusion relationship. Based on the local surface update results, a consistency check is performed after writing back. The updated surface segment is regenerated with the projection contour region, projection center trajectory and projection bandwidth distribution, and compared with the contour topology phase band set to obtain the set of deviation results after writing back. Based on the set of deviation results after writeback, perform writeback termination determination and generate a set of writeback termination results; The write-back termination determination includes performing write-back and holding on surface segments whose center deviation, bandwidth deviation, direction deviation, curvature deviation, and state deviation results are all lower than the corresponding thresholds; performing write-back again on surface segments that are still higher than the corresponding thresholds; and performing stabilization and freezing on surface segments that have reached the preset upper limit of write-back times. Based on the write-back termination result set, all surface fragments that have completed write-back retention, achieved the target after another write-back, and stabilized and frozen are uniformly encapsulated, and the output includes updated vertex positions, updated surface fragments, updated fragment affiliation relationships, and updated phase sources.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a character modeling scenario in an animation content production organization, using a single animation character illustration under the same art style as the input source. In actual production, designers often encounter the following problem: although the 3D character generated from the single illustration can form a basic shape, the outer edge of the hair, the edge of the clothing hem, the neckline connection, and the occlusion layer are prone to deviation, resulting in an inconsistency between the generated result and the original drawing in terms of outline and texture. This necessitates repeated manual adjustments, retouching, and layer smoothing, leading to a long production cycle and frequent rework. In this scenario, this invention first performs foreground separation, image normalization, line art enhancement, and contour distance transformation on the single animation character image to obtain a normalized character image. Then, it extracts the contour centerline, bandwidth, direction, curvature, open / closed state, and attachment state to form a contour topology phase band set. Furthermore, it establishes a contour topology map containing continuation, attachment, wrapping, and occlusion relationships, enabling the clear expression of the connection relationships between the character's hair, face, clothing, and auxiliary structures.

[0031] In practical applications, after the artists import the character illustrations into the modeling process, the phase-separated geometry generation network performs attachment, expansion, and wrapping generation on different contour regions based on the contour topology map, first obtaining the initial 3D character geometric model. Then, through projection mapping, it compares the model segment by segment with the contour topology phase band set to form the topology phase band deviation result. Based on this, local surface rewriting is performed on local areas with large deviations. During implementation, modeling logs, rewriting logs, and manual adjustment records are continuously recorded. Data items include contour center deviation records, bandwidth consistency records, direction deviation records, curvature deviation records, state deviation records, local surface rewriting trigger records, and manual modeling round records. On-site data shows that in the same batch of character modeling, the outer contour boundary of the character is closer to the original drawing, the wrapping layer between the hair and the face is more stable, the connection transition of the attached areas such as the hem and cuffs is more natural, the frequency of manual edge patching and manual point pulling is significantly reduced, the number of rework records continues to decrease, and the output model maintains a high degree of consistency in terms of contour fidelity, structural layer accuracy and geometric stability. This indicates that the present invention can effectively solve the problems of contour distortion, layer misalignment and local structural collapse in the 3D reconstruction of a single animation character image.

[0032] Table 1. Comprehensive Comparison of 3D Reconstruction Effects of Single Anime Characters

[0033] Overall, the invention demonstrates the most significant improvement in core metrics directly related to contour structure. The average deviation of the contour center decreased from 5.8 pixels to 3.6 pixels, the average bandwidth deviation from 4.9 pixels to 3.1 pixels, the average orientation deviation angle from 12.7 degrees to 8.4 degrees, and the average curvature deviation from 0.084 to 0.053. This indicates that the generated 3D geometric model more closely resembles the original anime character image in terms of boundary position, contour thickness, local orientation, and bending variations. This is because the reconstruction process does not rely solely on 2D appearance features to directly regress the 3D shape. Instead, it first extracts the contour centerline, bandwidth, orientation, curvature, open / closed state, and attachment state, then forms a contour topology phase band set, and further establishes a contour topology map. This allows contour structure information to be explicitly utilized during the geometry generation stage, effectively reducing the offset of hair tips, hemlines, necklines, and the boundaries of auxiliary structures.

[0034] In terms of hierarchical and structural relationships, this invention also demonstrates higher stability. The consistency rate for determining open / closed states increased from 88.6% to 95.2%, the consistency rate for determining attached states increased from 84.9% to 93.1%, the consistency rate for recognizing wrapping relationships increased from 82.4% to 91.3%, and the consistency rate for recognizing occlusion relationships increased from 80.8% to 90.5%. This indicates that the connections, enclosures, and occlusions between the character's hair and face, clothing and limbs, and accessories and the main body are expressed more accurately. While conventional methods can restore the basic volume of a character, they often rely on post-processing adjustments for the structural organization between local contours, which can easily lead to hierarchical misalignment and misjudgment of relationships. This invention, however, constrains contour segments through continuation, attachment, wrapping, and occlusion relationships, and then performs differentiated geometric generation processing according to the attachment phase, outward expansion phase, and wrapping phase, thus making it easier to maintain structural consistency in anime character scenes with complex edge hierarchies.

[0035] From the perspective of subsequent refinement and production efficiency, the advantages of this invention are also quite direct. After local surface rewriting, the residual deviation decreased from 4.2 to 2.5; the number of manual rewriting rounds per character decreased from 5.1 to 2.9; the modeling completion time per character was shortened from 46.8 minutes to 31.4 minutes; and the first-time approval rate increased from 72.6% to 89.4%. This data demonstrates that this invention not only improves the initial modeling quality but also reduces the workload of manual edge patching, repeated point pulling, and local reshaping by driving local surface rewriting through topological phase band deviation results. Especially for hair tips, skirt hem extension boundaries, cuff attachment areas, and occlusion boundaries, this invention can promptly identify deviations in band center, bandwidth, direction, curvature, and state after projection comparison and perform targeted rewriting, thereby reducing the frequency of repeated modifications after the model enters the manual rewriting stage.

[0036] The final model consistency score improved from 81.7 to 91.2, indicating that the present invention achieved a better balance between overall visual consistency, contour fidelity, and local structural stability. This improvement did not rely on optimization of a single metric, but rather stemmed from the continuous synergistic effect of input normalization, contour topology phase band extraction, contour topology map construction, phase-separated geometry generation, topology phase band deviation comparison, and local surface rewriting. Since the intermediate results of each stage can be reused in subsequent stages, contour structure information is continuously transmitted throughout the entire reconstruction process, thus avoiding the common problem of disconnect between front-end feature extraction and back-end geometry generation in conventional methods.

[0037] In summary, this invention can more accurately maintain the consistency of the character's outline center, bandwidth distribution, direction changes, curvature changes, and open / closed and attached states under the condition of a single anime character image. At the same time, it improves the accuracy of determining the wrapping and occlusion relationships, reduces the number of manual retouching rounds and the modeling completion time, and increases the first-time approval rate and the consistency of the final model. Thus, in the reconstruction of 3D anime characters, it takes into account the integrity of the outline, the accuracy of the structural hierarchy, the stability of local geometry, and the improvement of production efficiency.

[0038] The above are merely preferred embodiments 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 reconstructing 3D animation characters based on deep learning, characterized in that, Includes the following steps: The process involves acquiring a single anime character image, performing foreground separation, image normalization, line art enhancement, and contour distance transformation to generate a normalized character image. The normalized character image is input into the contour topology phase band extraction network to extract the contour centerline, bandwidth, direction, curvature, opening and closing state and attachment state, and generate a contour topology phase band set. A contour topology map containing continuation, attachment, wrapping, and occlusion relationships is established based on the contour topology phase band set. The normalized character image and contour topology map are input into the phase-separated geometry generation network to generate an initial 3D character geometric model according to the attached phase, the outward phase, and the wrapping phase. Perform projection mapping on the initial 3D character geometry model, compare it with the contour topology phase band set in terms of band center, bandwidth, direction, curvature and state correspondence, and generate topology phase band deviation results; Based on the topological phase band deviation results, a local surface rewrite is performed on the initial 3D character geometric model to obtain the 3D animation character geometric model.

2. The method for reconstructing 3D animation characters based on deep learning according to claim 1, characterized in that, The generation of the standardized character image specifically includes: The process involves acquiring a single anime character image, performing pixel validity checks, alpha channel parsing, and color space unification to generate the original character image. Perform foreground separation on the original character image, output a foreground probability map, perform consistency fusion with the transparent region marker map or the background candidate region map, calculate the fused foreground response value, and obtain the foreground mask map; Based on the foreground mask, the minimum bounding region of the foreground is extracted. The original character image is then cropped, centered, translated, and scaled to obtain a cropped character image. Boundary padding is then performed based on the aspect ratio of the minimum bounding region of the foreground and the preset target side length to generate a scaled character image. Perform image normalization on scale-uniform character images to generate normalized image results; The image normalization result is input into the line art enhancement network to extract the character's outer contour response, local boundary response, and high curvature inflection point response. The background region response is suppressed based on the foreground mask to generate the line art enhancement image. Based on the enhanced line art image, closed and open contour regions are extracted. Contour distance transformation is performed to obtain inner and outer contour distance maps, respectively. These maps are then encapsulated with the image normalization result, foreground mask map, and enhanced line art image in channel order to generate a normalized character image.

3. The method for reconstructing 3D animation characters based on deep learning according to claim 1, characterized in that, The generation of the contour topology phase band set specifically includes: Normalized character images are input into a contour topology phase band extraction network in channel order. Joint encoding is performed on the normalized character images to extract multi-scale contour response features, region boundary features, and distance distribution features, generating a contour extraction input feature map. Perform centerline extraction processing on the contour extraction input feature map to generate contour centerline results; Based on the contour centerline results and the contour extraction input feature map, bandwidth extraction processing is performed to generate bandwidth results that correspond one-to-one with the contour centerline. Based on the contour centerline results, bandwidth results, inner contour distance map and outer contour distance map, perform direction and curvature extraction processing to generate contour geometric description results; Based on the contour geometry description results and the foreground mask, state determination processing is performed to obtain the opening / closing state and attachment state corresponding to each contour centerline, and generate contour state results. The contour centerline results, bandwidth results, contour geometric description results, and contour state results are associated and encapsulated by contour segments. Adjacent contour segments with the same attributes are merged, and contour segments with conflicting attributes are segmented by boundary, generating a contour topology phase band set.

4. The method for reconstructing 3D animation characters based on deep learning according to claim 3, characterized in that, The centerline extraction process includes refining and clustering the continuous contour response region and suppressing breakpoints.

5. The method for reconstructing 3D animation characters based on deep learning according to claim 1, characterized in that, The generation of the contour topology map specifically includes: Read the contour topology phase band set, establish a contour segment index according to the contour center line, bandwidth result, contour geometric description result and contour state result corresponding to each contour topology phase band, and generate a contour segment candidate association set based on the positional continuity of the contour center line in the normalized character image and the segment affiliation relationship. Based on the contour segment index and the candidate association set of contour segments, a continuation relationship determination is performed to obtain the continuation relationship set; Based on the contour fragment index, foreground mask map, and contour state results, an attachment relationship determination is performed to obtain an attachment relationship set. Based on the contour segment index, inner contour distance map, outer contour distance map and contour geometric description results, the wrapping relationship determination is performed to obtain the wrapping relationship set; Occlusion relationship determination is performed based on the contour fragment index, line art enhancement image, foreground mask image and contour state results to obtain an occlusion relationship set; Write the sets of extension relations, attachment relations, wrapping relations, and occlusion relations into the relation index table according to the relation type. Perform relation conflict adjudication on the contour fragments that simultaneously satisfy multiple relations to obtain the relation adjudication result set. Using the contour topology phase band set as the node set, and the continuation relationship, attachment relationship, wrapping relationship and occlusion relationship in the relationship adjudication result set as the edge set, the contour topology graph is generated by associating and encapsulating the edges according to edge type, edge start and end segments and corresponding position indexes.

6. The method for reconstructing 3D animation characters based on deep learning according to claim 1, characterized in that, The generation of the initial 3D character geometric model specifically includes: Based on the phase-splitting geometric generation network, the network extracts the overall appearance features, local texture features, and spatial distribution features of the character from the normalized character image. It also extracts the node attribute features, edge relationship features, and relationship transfer features from the contour topology graph. The network is then aligned and fused according to the position index relationship in the contour topology phase band set to generate a geometric generation input feature set. Based on the geometrically generated input feature set, phase determination is performed on each contour topology phase band in the contour topology map to generate a phase annotation result set; Based on the phase annotation result set and the geometric generation input feature set, a phase sub-generation unit is established to obtain the attached phase local surface result, the expanded phase local surface result, and the wrapped phase local surface result. The attached phase local surface results, the extended phase local surface results, and the wrapped phase local surface results are processed according to the set of continuation relationships, attachment relationships, wrapping relationships, and occlusion relationships in the contour topology map. Adjacency stitching, boundary smoothing, and hierarchical adjudication are performed to generate the phase-separated surface stitching results. Based on the phase-separated surface splicing results, the basic character mesh skeleton is established. Continuous splicing is performed on local surfaces with a continuation relationship, connection and fixation are performed on local surfaces with an attachment relationship, enclosing and nesting are performed on local surfaces with a wrapping relationship, and front-to-back hierarchical sorting and occlusion preservation are performed on local surfaces with an occlusion relationship to generate the initial character mesh result. Based on the initial character mesh results and the image normalization results in the normalized character image, region uniformity processing is performed. Mesh subdivision density allocation, boundary orientation straightening, and local curvature buffering are performed on the character outer contour region, local boundary region, and auxiliary structure region, respectively, to generate geometrically uniform results. The geometrically consistent results are associated and encapsulated according to vertex position, surface segment, segment affiliation, and phase source to obtain the initial 3D character geometric model.

7. The method for reconstructing 3D animation characters based on deep learning according to claim 6, characterized in that, The phase generation unit performs attachment generation, expansion generation, and encirclement generation processes on the attached phase, the outward expansion phase, and the enclosed phase, respectively.

8. The method for reconstructing 3D animation characters based on deep learning according to claim 1, characterized in that, The generation of the topological phase band deviation result specifically includes: Read the initial 3D character geometry model, establish a projection input set based on vertex position, surface segment, segment affiliation relationship and phase source, determine the viewpoint alignment parameters by combining the position index relationship in the normalized character image, and generate a projection input alignment set; Perform projection mapping on the projection input alignment set, and generate the projection contour region, projection center trajectory and projection bandwidth distribution based on the surface fragment boundaries, phase source and front and back hierarchy order in the initial 3D character geometry model, to obtain the projection topology phase band set; The projected topological phase band set is registered with the contour topological phase band set to generate a set of topological phase band alignment results. Based on the topological phase band alignment result set, band center comparison is performed, and the center trajectory position deviation, extension length deviation, and endpoint connectivity deviation of each corresponding segment are calculated segment by segment to obtain the band center deviation result; Based on the topological phase band alignment result set, bandwidth comparison, direction comparison and curvature comparison are performed. The differences in the inner and outer width distribution, local orientation difference and bending change difference of each corresponding segment are calculated segment by segment to obtain the bandwidth deviation result, direction deviation result and curvature deviation result. Based on the set of topological phase band alignment results, a state comparison is performed to determine the consistency of the opening and closing states, the consistency of the attachment states, and the consistency of the continuation, attachment, wrapping, and occlusion relationships inherited from the contour topology graph of each corresponding segment, and to obtain the state deviation results. The center deviation results, bandwidth deviation results, direction deviation results, curvature deviation results, and state deviation results are associated and summarized according to the contour segments. Deviation propagation and integration are performed on continuous segments, and relationship priority adjudication is performed on conflicting segments to generate a segment-level deviation set. The fragment-level deviation set is backfilled and encapsulated according to the contour centerline, bandwidth result, contour geometric description result and contour state result in the contour topology phase band set to obtain the topology phase band deviation result.

9. A method for reconstructing 3D animation characters based on deep learning according to claim 8, characterized in that, The registration is performed according to the position index relationship, the segment belonging relationship and the phase source. For the projection segments that have not formed a corresponding relationship, a neighborhood search is performed to supplement the registration. For the projection segments with multiple candidate correspondences, the minimum position deviation is selected.

10. A method for reconstructing 3D animation characters based on deep learning according to claim 1, characterized in that, The generation of the geometric model of the 3D animation character specifically includes: Read the initial 3D character geometric model and topological phase band deviation results, establish a local surface write-back input set according to the position index relationship of contour segments, segment belonging relationship and phase source, and collect the deviations of each surface segment based on the center deviation result, bandwidth deviation result, direction deviation result, curvature deviation result and state deviation result to generate a local surface write-back candidate set; Based on the local surface writeback candidate set, perform writeback trigger determination and generate a writeback trigger result set; Establish a sub-action write-back result set based on the write-back trigger result set and the phase source; Write the set of sub-action writeback results back to the corresponding surface fragment in the initial 3D character geometric model to generate local surface update results; Based on the local surface update results, a consistency check is performed after writing back. The updated surface segment is regenerated with the projection contour region, projection center trajectory and projection bandwidth distribution, and compared with the contour topology phase band set to obtain the set of deviation results after writing back. Based on the set of deviation results after writeback, perform writeback termination determination and generate a set of writeback termination results; Based on the write-back termination result set, all surface fragments that have completed write-back retention, achieved the target after another write-back, and stabilized and frozen are uniformly encapsulated, and the output includes updated vertex positions, updated surface fragments, updated fragment affiliation relationships, and updated phase sources.

Images