A visual intelligence-based picture labeling information recognition method

Abstract

The application discloses a picture mark information recognition method based on visual intelligence, comprising the following steps: obtaining and preprocessing a picture to be recognized; extracting multi-source visual trace features; recognizing mark superimposed candidate pixel regions in the picture to be recognized; performing mark layer de-coherence separation processing on the picture to be recognized, generating an original image content layer and a mark information candidate layer and recording a layer separation mapping relationship; performing text, line segment and contour recognition on the mark information candidate layer; projecting a mark element set back to the original image content layer and performing mark element back projection adaptive correction; determining a corresponding relationship between the mark element set and a marked object and generating a picture mark information recognition result. The application utilizes multi-source visual trace analysis and layer de-coherence separation methods, realizes accurate association between picture mark information and a marked object, and has the advantages of high recognition accuracy and strong complex scene adaptability.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, and in particular to a method for recognizing image annotation information based on visual intelligence. Background Technology

[0002] With the development of image processing and visual recognition technologies, image annotation information recognition is widely used in scenarios such as industrial inspection, medical image analysis, engineering drawing interpretation, and educational image processing. Existing technologies typically employ OCR recognition, edge detection, or object detection methods to identify text, arrows, borders, and selected content in images, and then parse the annotation content based on spatial relationships to extract image annotation information.

[0003] Existing technologies are susceptible to interference from transparent overlays, texture occlusion, compression residual differences, and complex backgrounds when adding annotations and overlaying them onto the main image content. This makes it difficult to accurately separate the annotation area from the main image content. Most existing technologies determine the annotation objects directly based on two-dimensional positional relationships, lacking correction processing for the correspondence between the annotation layer and the original image content. This can easily lead to problems such as arrow endpoint offsets, misaligned selection areas, and incorrect association of annotation objects, affecting the accuracy of image annotation information recognition.

[0004] Therefore, how to provide a method for recognizing image annotation information based on visual intelligence 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 visual intelligence-based image annotation information recognition method. This invention utilizes multi-source visual trace analysis and layer decoherence separation to achieve accurate association between image annotation information and the annotated object, and has the advantages of high recognition accuracy and strong adaptability to complex scenes.

[0006] A method for recognizing image annotation information based on visual intelligence according to an embodiment of the present invention includes the following steps: Acquire and preprocess the image to be recognized to generate basic visual data for the image; Extracting multi-source visual trace features based on image-based visual data; Based on the multi-source visual trace features, identify the candidate pixel regions with superimposed annotations in the image to be identified, and generate a set of superimposed annotation regions with annotation type and region attributes; The set of superimposed labeled regions is input into the label layer segmentation model based on SegFormer. The label layer decoherence separation process is performed on the image to be recognized to generate the original image content layer and the label information candidate layer and record the layer separation mapping relationship. The candidate layer of annotation information is subjected to text, line segment and contour recognition to generate a set of annotation elements; According to the layer separation mapping relationship, the set of labeled elements is back-projected to the original image content layer, and adaptive correction of the label element back-projection is performed based on the object boundary response in the original image content layer to generate the label element back-projection result; Based on the back projection results of the labeled elements and the object regions in the original image content layer, the correspondence between the set of labeled elements and the labeled objects is determined, and the image annotation information recognition result is generated. Optionally, the preprocessing includes size normalization, color space conversion, noise suppression, and image enhancement.

[0007] Optionally, the extraction of the multi-source visual trace features specifically includes: HSV color space data and Lab color space data are extracted from basic visual data of images. HSV color space features are generated based on the hue distribution, saturation distribution and lightness distribution in the HSV color space data. Lab color space features are generated based on the lightness component distribution and chromaticity component distribution in the Lab color space data. Color change detection is performed on adjacent pixel regions in the basic visual data of the image. Based on the color change amplitude of adjacent pixel regions in the HSV color space and Lab color space, color channel abrupt change features are generated. Laplacian edge detection is performed on the basic visual data of the image to extract the line width distribution, continuous line length and line direction variation in the edge response region. Line width consistency features are generated based on the consistent state of the line width distribution, and edge closure morphology features are generated based on the closed contour distribution in the edge response region. We perform texture continuity analysis on local texture regions in basic visual data of images, and generate local texture interruption features based on texture direction interruption locations, texture gray-level abrupt change locations, and texture repetition pattern destruction locations. Based on the local brightness superposition state, background transparency state, and JPEG compression block residual distribution in the basic visual data of the image, transparent superposition trace features and JPEG compression residual difference features are generated. The transparent superposition trace features, JPEG compression residual difference features, HSV color space features, Lab color space features, color channel abrupt change features, line width consistency features, edge closure shape features, and local texture interruption features are combined into multi-source visual trace features.

[0008] Optionally, the generation of the set of labeled overlay regions specifically includes: The multi-source visual trace features are subjected to coordinate alignment and scale unification processing, and the processed multi-source visual trace features are mapped to the corresponding pixel regions of the image to be identified to generate a multi-source visual trace response map. Calculate the confidence score of the superimposed annotations for each pixel region in the image to be identified based on the multi-source visual trace response map; Candidate pixel regions for annotation overlay are selected based on the annotation overlay confidence score and the annotation overlay judgment threshold. Connectivity aggregation and boundary correction are then performed on the candidate pixel regions for annotation overlay to generate candidate annotation regions. Based on the text response, line segment direction, endpoint shape, closed boundary, number character response, and color distribution characteristics of the candidate annotation area, the candidate annotation area is identified to obtain the annotation type area. The annotation type includes text annotation, arrow annotation, leader line annotation, box selection annotation, circle selection annotation, numbered annotation, and color label. Perform region deduplication, region merging, and region attribute configuration on candidate annotation regions corresponding to different annotation types to generate a set of annotation overlay regions.

[0009] Optionally, the records that record the layer separation mapping relationship specifically include: A mask for the superimposed annotation regions is generated based on the set of superimposed annotation regions. The mask for the superimposed annotation regions is then fused with the basic visual data of the image to generate input data for annotation layer segmentation. The annotation layer segmentation input data is input into the SegFormer-based annotation layer segmentation model. The hierarchical Transformer encoder performs multi-stage feature extraction on the annotation layer segmentation input data to generate multi-scale visual features at different resolutions. Multi-scale visual features of different resolutions are input into a multilayer perceptron decoder for pixel-level layer classification, generating layer segmentation results. Based on the layer segmentation results, the pixels in the image to be identified are classified into layers to generate the original image content layer and the annotation information candidate layer. Boundary consistency correction is performed on the original image content layer and the annotation information candidate layer, and the pixel coordinate correspondence between the original image content layer and the annotation information candidate layer is recorded to generate the layer separation mapping relationship.

[0010] Optionally, the generation of the set of labeled elements specifically includes: Text region detection is performed on the candidate layer of annotation information. Character segmentation, character contour feature extraction and character recognition are performed on the detected text regions based on OCR to obtain the annotation text and number in the candidate layer of annotation information and obtain the corresponding text box position and text box boundary. Edge extraction and binarization are performed on the candidate layer of annotation information to generate a linear annotation edge map. Candidate annotation line segments are detected from the linear annotation edge map based on probabilistic Hough transform, and the endpoints, lengths and directions of the candidate annotation line segments are extracted. Perform endpoint structure analysis and connection relationship analysis on candidate annotation line segments. Configure line segments with arrow endpoint structures as arrow line segments, configure line segments connecting the annotation text and the direction of the main content of the image as leader line segments, and configure line segments that connect with each other and form a rectangular boundary as frame line segments. Contour detection is performed on the candidate layer of annotation information to extract closed contours and color-connected contours. Contours with curved enclosing shapes in closed contours are configured as selection lines, and contours with continuous color distribution and regional filling features are configured as color markers. The annotation text, number, arrow segments, leader segments, frame segments, selection lines, and color identifiers are merged to generate a set of annotation elements. Based on the text box position, text box boundary, line segment endpoint coordinates, contour boundary coordinates, and identification category of each annotation element, the element position, element boundary, and element type are configured for each annotation element in the annotation element set.

[0011] Optionally, the generation of the back projection results of the labeled elements specifically includes: Read the layer separation mapping relationship, map the element positions and element boundaries in the annotation element set from the annotation information candidate layer to the original image content layer, and generate the initial back projection position of the annotation text, the initial back projection position of the number, the initial endpoint position of the arrow segment, the initial endpoint position of the leader segment, the initial coverage area of ​​the frame line, and the initial coverage area of ​​the circle line. Object boundary response is extracted from the content layer of the original image to generate an object boundary response map; Based on the object boundary response map, the initial endpoint positions of the arrow segment and the initial endpoint positions of the leader segment are corrected to obtain the back projection positions of the arrow segment endpoints and the back projection positions of the leader segment endpoints. Based on the object boundary response map, boundary snapping correction is performed on the initial coverage area of ​​the frame line and the initial coverage area of ​​the delineation line to obtain the frame line coverage area and the delineation coverage area. Based on the initial back projection position of the annotation text, the initial back projection position of the number, and the object boundary response map, the location correction of the neighboring object is performed to obtain the back projection position of the annotation text and the back projection position of the number. The results of the back projection positions of the annotation text, the number, the arrow segment endpoints, the leader segment endpoints, the frame coverage area, and the circled coverage area are then encapsulated to generate the annotation element back projection result.

[0012] Optionally, the generation of the image annotation information recognition result specifically includes: The object region is extracted from the content layer of the original image to obtain a set of object regions, and the object region position and object region boundary are configured for each object region in the set of object regions. Based on the back projection results of the labeled elements and the set of object regions, candidate relationships of labeled objects are constructed, and the object association value corresponding to each candidate relationship of labeled objects is calculated; For each annotation element in the annotation element set, the object region with the largest object association value is determined as the annotated object, and the annotation element identifier and the annotated object identifier are written into the same annotation object relationship item to obtain the generated correspondence; The image annotation information recognition result is generated by associating the annotation element set, the annotated object, and the corresponding relationship. The image annotation information recognition result includes the annotation content, annotation type, annotation layer position, annotated object, the position of the annotated object, and the correspondence between the annotated objects.

[0013] The beneficial effects of this invention are: This invention constructs multi-source visual trace features by jointly analyzing color differences, line width consistency, closed edge morphology, texture continuity, transparent overlay traces, and compression residual differences in the image to be identified. Based on these features, candidate pixel regions for annotation overlay are identified, enabling accurate localization of subsequently added annotations in complex backgrounds, semi-transparent overlays, and texture-occluded scenes. Compared to existing technologies that rely solely on text detection or simple edge recognition, this invention effectively distinguishes visual differences between the main image content and subsequently added annotations, reducing interference from background textures, compression noise, and color mixing on the annotation region recognition results, thus improving the accuracy of annotation overlay region recognition and adaptability to complex scenes.

[0014] This invention inputs the set of superimposed labeled regions into a SegFormer-based label layer segmentation model, performs decoherence separation processing on the labeled layers of the image to be identified, and backprojects the set of labeled elements onto the original image content layer based on the layer separation mapping relationship. This establishes a stable pixel coordinate correspondence between the candidate layer of labeled information and the main content of the image. Through object boundary response extraction, endpoint repositioning correction, boundary snapping correction, and neighboring object positioning correction, adaptive correction can be performed on the endpoints of arrow segments, leader segments, bounding box coverage areas, circled coverage areas, and the position of labeled text. This reduces the boundary offset and positional error caused by layer separation, improves the positional consistency between labeled elements and the main content of the image, and avoids problems such as misaligned labeled objects, drifting of selected areas, and deviation of arrow pointing.

[0015] This invention constructs candidate relationships for labeled objects based on the back projection results of labeled elements and the set of object regions. It calculates object association values ​​according to different relationship types, enabling the establishment of corresponding object association mechanisms for text labels, numbered labels, arrow labels, leader line labels, box selection labels, and circle selection labels. By determining the object region with the highest object association value as the labeled object and generating a correspondence between labeled elements and labeled objects, it achieves precise association between image labeling information and the main content of the image. This improves the accuracy and completeness of image labeling information recognition results, ensuring that the generated image labeling information recognition results accurately represent the labeling content, labeling type, labeling layer position, and the association relationships between labeled objects. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a visual intelligence-based image annotation information recognition method proposed in this invention; Figure 2 This is a flowchart of the decoherence separation process for the image annotation information recognition method based on visual intelligence proposed in this invention. Figure 3 This is a flowchart illustrating the generation of the backprojection results of the labeled elements in a visual intelligence-based image annotation information recognition method proposed in this invention. Detailed Implementation

[0017] 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.

[0018] refer to Figures 1-3 A method for recognizing image annotation information based on visual intelligence includes the following steps: Acquire and preprocess the image to be recognized to generate basic visual data for the image; Extracting multi-source visual trace features based on image-based visual data; Based on the multi-source visual trace features, identify the candidate pixel regions with superimposed annotations in the image to be identified, and generate a set of superimposed annotation regions with annotation type and region attributes; The set of superimposed labeled regions is input into the label layer segmentation model based on SegFormer. The label layer decoherence separation process is performed on the image to be recognized to generate the original image content layer and the label information candidate layer and record the layer separation mapping relationship. The candidate layer of annotation information is subjected to text, line segment and contour recognition to generate a set of annotation elements; According to the layer separation mapping relationship, the set of labeled elements is back-projected to the original image content layer, and adaptive correction of the label element back-projection is performed based on the object boundary response in the original image content layer to generate the label element back-projection result; Based on the back projection results of the labeled elements and the object regions in the original image content layer, the correspondence between the set of labeled elements and the labeled objects is determined, and the image labeling information recognition results are generated.

[0019] In this embodiment, preprocessing includes size normalization, color space conversion, noise suppression, and image enhancement. Size normalization adjusts the image to be recognized to a uniform input size while preserving the relative positional relationship between the labeled elements and the main content of the image. Color space conversion converts the image to be recognized to the HSV and Lab color spaces to extract color and brightness differences. Noise suppression reduces acquisition and compression noise in the image. Image enhancement enhances the visual difference between the added labeled content and the main content of the image.

[0020] In this embodiment, the extraction of multi-source visual trace features specifically includes: HSV color space data and Lab color space data are extracted from basic visual data of images. HSV color space features are generated based on the hue distribution, saturation distribution and lightness distribution in the HSV color space data. Lab color space features are generated based on the lightness component distribution and chromaticity component distribution in the Lab color space data. Color change detection is performed on adjacent pixel regions in the basic visual data of the image. Based on the color change amplitude of adjacent pixel regions in the HSV color space and Lab color space, color channel abrupt change features are generated. The color channel abrupt change features include hue abrupt change position, saturation abrupt change position, brightness abrupt change position, and chroma abrupt change position. Laplacian edge detection is performed on the basic visual data of the image. Laplacian edge detection includes grayscale conversion of the basic visual data, calculation of the second-order gradient response of each pixel in the grayscale image based on the Laplacian operator, determination of edge response intensity based on the second-order gradient response, selection of edge response pixels according to edge response intensity, and connection component analysis of the edge response pixels to obtain edge response regions. Line width distribution, continuous line length, and line direction variation are extracted from the edge response regions. Line width consistency features are generated based on the consistent state of the line width distribution, and edge closure morphology features are generated based on the distribution of closed contours in the edge response regions. Line width consistency features include the width variation state within the same line region and the width similarity state between adjacent line regions. Edge closure morphology features include rectangular closed boundaries, circular closed boundaries, and irregularly selected closed boundaries. Texture continuity analysis is performed on local texture regions in basic visual data of images. Based on the location of texture direction interruption, texture grayscale abrupt change, and texture repetition pattern destruction, local texture interruption features are generated. Local texture interruption features are used to characterize the texture extension interruption state formed after adding labeled content to cover the main content of the image. The texture extension interruption state includes texture direction being truncated, texture grayscale continuity being destroyed, and texture repetition pattern being occluded. Based on the local brightness overlay state, background transparency state, and JPEG compression block residual distribution in the basic visual data of the image, transparent overlay trace features and JPEG compression residual difference features are generated. The transparent overlay trace features, JPEG compression residual difference features, HSV color space features, Lab color space features, color channel abrupt change features, linewidth consistency features, edge closure shape features, and local texture interruption features are combined into multi-source visual trace features. The transparent overlay trace features are used to characterize the brightness overlay state and background transparency state formed after the semi-transparent annotation content covers the main content of the image. The JPEG compression residual difference features are used to characterize the difference in compression residual distribution between the area where the added annotation content is located and the area where the main content of the image is located.

[0021] In this embodiment, the generation of the overlay region set specifically includes: The multi-source visual trace features are subjected to coordinate alignment and scale unification processing, and the processed multi-source visual trace features are mapped to the corresponding pixel regions of the image to be identified to generate a multi-source visual trace response map. The confidence scores of annotation overlay for each pixel region in the image to be identified are calculated based on the multi-source visual trace response map. The confidence scores are obtained by weighted summation after normalization of color difference response, line width consistency response, closed edge response, texture interruption response, transparency overlay response, and compression residual response. Color difference response is used to characterize the degree of color channel abrupt change in the image to be identified, line width consistency response is used to characterize the degree of consistency of line width distribution, closed edge response is used to characterize the integrity of closed contour boundaries, texture interruption response is used to characterize the degree of texture continuity disruption, transparency overlay response is used to characterize the local brightness overlay state and background transparency state, and compression residual response is used to characterize the difference in compression residual distribution between local areas and surrounding areas. Candidate pixel regions for annotation overlay are selected based on the annotation overlay confidence score and the annotation overlay judgment threshold. Connectivity aggregation and boundary correction are then performed on these candidate pixel regions to generate candidate annotation regions. Specifically: pixel regions with an annotation overlay confidence score greater than or equal to the annotation overlay judgment threshold are identified as candidate pixel regions for annotation overlay; adjacent candidate pixel regions with pixel connectivity are grouped into the same candidate region unit; candidate region units with an area smaller than a preset area threshold are deleted; and the remaining candidate region units undergo circumscribed boundary extraction, boundary gap filling, and boundary smoothing to generate candidate annotation regions. The annotation overlay judgment threshold is obtained by summing the average annotation overlay confidence score and the standard deviation of the annotation overlay confidence score for all pixel regions; the preset area threshold is obtained by multiplying the total pixel area of ​​the image to be identified by the minimum annotation area ratio. Based on the text response, line segment direction, endpoint shape, closed boundary, number character response, and color distribution characteristics of the candidate annotation area, the candidate annotation area is identified to obtain the annotation type area. The annotation types include text annotation, arrow annotation, leader line annotation, box selection annotation, circle selection annotation, numbered annotation, and color label. Type recognition includes: identifying candidate annotation areas with continuous text responses and consistent character arrangement directions as text annotations; identifying candidate annotation areas with a combination of numeric and alphanumeric characters in the numbered character responses as numbered annotations; identifying candidate annotation areas with arrow endpoints and line segments with a unidirectional pointing structure as arrow annotations; identifying candidate annotation areas containing only continuous line segments and without arrow endpoints as leader line annotations; identifying candidate annotation areas with right angles between the long and short sides of closed boundaries as bounding box annotations; identifying candidate annotation areas with closed boundaries composed of continuous curves as circle annotations; and identifying candidate annotation areas with continuous color distribution and the same color pixels accounting for more than half of the total number of pixels in the candidate annotation area as color identifiers. For candidate annotation regions corresponding to different annotation types, perform region deduplication, region merging, and region attribute configuration to generate a set of annotation overlay regions. The region attributes include region location, region boundary, region type, and annotation overlay confidence.

[0022] In this embodiment, the recording of layer separation mapping relationships specifically includes: An annotation overlay region mask is generated based on the set of annotated overlay regions. This mask is then fused with the basic visual data of the image to generate the input data for annotation layer segmentation. Specifically, the region contours are extracted from the region boundaries in the set of annotated overlay regions. Pixels within the region contours are marked, with pixels belonging to the set designated as annotated pixels and pixels not belonging to the set designated as background pixels, generating an annotation overlay region mask with the same size as the image to be identified. This mask is then used as a mask channel and concatenated with the image channels in the basic visual data of the image to generate the input data for annotation layer segmentation. The annotation layer segmentation input data is fed into a SegFormer-based annotation layer segmentation model. A hierarchical Transformer encoder performs multi-stage feature extraction on the annotation layer segmentation input data to generate multi-scale visual features at different resolutions. Specifically, the annotation layer segmentation input data is divided into multiple image patch sequences, the image patch sequences are positionally encoded and fed into a hierarchical Transformer encoder; self-attention feature calculation and feature downsampling are performed on the image patch sequences through different stages of Transformer feature extraction layers to generate feature maps corresponding to different resolutions; annotation boundary features are generated based on the edge response intensity and edge continuity distribution in the feature maps; background texture features are generated based on the texture direction distribution and texture repetition distribution in the feature maps; annotation coverage features are generated based on the region brightness coverage and color coverage status in the feature maps; and layer boundary features are generated based on the pixel change distribution between the main content of the image and the subsequently added annotation content. Multi-scale visual features at different resolutions are input into a multilayer perceptron decoder for pixel-level layer classification to generate layer segmentation results. Specifically, the channel dimensions of the multi-scale visual features at different resolutions are unified, and the unified multi-scale visual features are mapped to the same resolution. Feature splicing and feature fusion are performed on the mapped multi-scale visual features to generate layer fusion features. Based on the layer fusion features, the layer classification probability of each pixel in the image to be identified is calculated, and the pixel category with the highest layer classification probability is determined as the layer classification category of the corresponding pixel, generating layer segmentation results. The layer classification category includes the pixel category of the original image content and the candidate pixel category of the annotation information. The calculation of layer attribution probability includes: inputting the feature vector corresponding to each pixel in the layer fusion features into the classification layer of the multilayer perceptron decoder, and weighting it with the original image content pixel category weight and the annotation information candidate pixel category weight in the classification layer to obtain the original image content pixel category score and the annotation information candidate pixel category score; normalizing the original image content pixel category score and the annotation information candidate pixel category score to obtain the layer attribution probability of each pixel belonging to the original image content pixel category and the layer attribution probability of each pixel belonging to the annotation information candidate pixel category; the original image content pixel category weight and the annotation information candidate pixel category weight are the classification layer weights obtained after training the annotation layer segmentation model; Based on the layer segmentation results, the pixels in the image to be recognized are classified into layers to generate an original image content layer and a candidate layer for annotation information. Specifically, the layer category corresponding to each pixel in the layer segmentation results is extracted. Pixels whose layer category is the original image content pixel category are assigned to the original image content layer, and pixels whose layer category is the candidate annotation information pixel category are assigned to the candidate annotation information layer. Neighborhood pixel completion processing is performed on missing pixel regions in the original image content layer, and isolated noise pixels in the candidate annotation information layer are removed to generate the original image content layer and the candidate annotation information layer. The original image content layer retains the main content of the image, and the candidate annotation information layer retains the main content and adds annotation content. Boundary consistency correction is performed on the original image content layer and the annotation information candidate layer, and the pixel coordinate correspondence between the original image content layer and the annotation information candidate layer is recorded to generate a layer separation mapping relationship; the pixel coordinate correspondence is used to maintain the consistency between the position of the annotation elements and the position of the main content of the image when the annotation element set is back-projected to the original image content layer. Boundary consistency correction includes: extracting the layer boundary edges between the original image content layer and the annotation information candidate layer; performing boundary filling processing on the boundary gap areas in the layer boundary edges; performing boundary smoothing processing on the jagged boundary areas; and removing isolated pixels near the boundary. Boundary gap areas are areas where there is a break between the endpoints of adjacent edges in the layer boundary edges; jagged boundary areas are areas where the extension direction of adjacent edge pixels in the layer boundary edges alternates, causing the layer boundary edges to form a discontinuous folded shape.

[0023] In this embodiment, the generation of the annotation element set specifically includes: Text region detection is performed on the candidate layer of annotation information. The detected text regions are then processed using OCR-based character segmentation, character contour feature extraction, and character recognition. Specifically: connected character region segmentation is performed on the text regions in the candidate layer to obtain multiple character regions; the character contour shape, character edge distribution, and character stroke structure corresponding to each character region are extracted to generate character contour features; character matching and character recognition are performed on each character region based on the character contour features to obtain the corresponding character content; and character combination is performed on the continuously arranged character content to obtain the annotation text and number in the candidate layer of annotation information, and the corresponding text box position and text box boundary are obtained. Edge extraction and binarization are performed on the candidate layer of annotation information to generate a linear annotation edge map. Candidate annotation line segments are detected from the linear annotation edge map based on probabilistic Hough transform, and the endpoints, lengths and directions of the candidate annotation line segments are extracted. The generation of the linear annotation edge map includes: performing grayscale conversion on the annotation information candidate layer, calculating the grayscale difference of each pixel in the horizontal and vertical directions respectively, and using the sum of the absolute values ​​of the horizontal and vertical grayscale differences as the gradient change intensity of the corresponding pixel; calculating the average gradient change intensity of all pixels in the annotation information candidate layer, and using the average gradient change intensity as the edge response threshold; identifying pixels with gradient change intensity greater than or equal to the edge response threshold as edge pixels, identifying pixels with gradient change intensity less than the edge response threshold as non-edge pixels, and generating a linear annotation edge map based on the spatial distribution of the edge pixels; The detection of candidate labeled line segments includes: randomly selecting edge pixels from the edge pixels in the linear labeled edge map, performing cumulative statistics on line parameters, calculating the number of edge pixel clusters corresponding to different line segment directions and positions, and determining the set of edge pixels with an edge pixel cluster number greater than the line segment detection threshold as candidate labeled line segments, where the line segment detection threshold is the sum of the average number of edge pixel clusters and the standard deviation of the number of edge pixel clusters in each edge pixel set; determining the line segment endpoints based on the starting and ending edge pixel positions in the candidate labeled line segments, determining the line segment length based on the pixel distance between the line segment endpoints, and determining the line segment direction based on the pixel extension direction between the line segment endpoints; The candidate annotation line segments are analyzed for endpoint structure and connectivity. Line segments with arrow endpoint structures are configured as arrow line segments, line segments connecting the annotation text and the main content of the image are configured as leader line segments, and line segments that connect to each other and form a rectangular boundary are configured as frame line segments. Arrow endpoint structure is defined as a candidate annotation line segment with two short line segments at the endpoints of the candidate line segment whose included angle is less than 90 degrees and converge toward the same endpoint. Connectivity analysis includes: determining whether the first endpoint of the candidate annotation line segment is located within the neighborhood of the text box corresponding to the annotation text or number, and determining whether the second endpoint of the candidate annotation line segment is located within the neighborhood of the area where the main content of the image is located. The neighborhood of the text box boundary is the area formed by extending the text box boundary outward by a preset pixel width, and the neighborhood of the area where the main content of the image is located is the area formed by extending the boundary of the area where the main content of the image is located outward by a preset pixel width. The preset pixel width is one-hundredth of the pixel length of the short side of the image to be identified. Contour detection is performed on the candidate layer of annotation information to extract closed contours and color-connected contours. Contours with curved enclosing shapes in closed contours are configured as selection lines, and contours with continuous color distribution and region filling characteristics are configured as color markers. Color distribution continuity means that the color difference between adjacent pixels within a color-connected contour is less than the color continuity threshold. The color difference is the sum of the absolute values ​​of the hue difference, saturation difference, and lightness difference of adjacent pixels in the HSV color space. The color continuity threshold is the average value of the color difference of all adjacent pixels within a color-connected contour. Region filling characteristics mean that the proportion of pixels of the same color within a color-connected contour is greater than one-half of the total number of pixels within the color-connected contour. The annotation text, number, arrow segments, leader segments, frame segments, selection lines, and color identifiers are merged to generate a set of annotation elements. Based on the text box position, text box boundary, line segment endpoint coordinates, contour boundary coordinates, and identification category of each annotation element, the element position, element boundary, and element type are configured for each annotation element in the annotation element set.

[0024] In this embodiment, the generation of the back projection result of the labeled elements specifically includes: Read the layer separation mapping relationship, map the element positions and element boundaries in the annotation element set from the annotation information candidate layer to the original image content layer, and generate the initial back projection position of the annotation text, the initial back projection position of the number, the initial endpoint position of the arrow segment, the initial endpoint position of the leader segment, the initial coverage area of ​​the frame line, and the initial coverage area of ​​the circle line. Specifically: Read the element position and boundary coordinates of each annotation element in the annotation element set, and query the original image content layer pixel coordinates corresponding to the element position and boundary coordinates in the layer separation mapping relationship; map the center coordinates of the text boxes corresponding to the annotation text and number to the original image content layer to generate the initial back projection position of the annotation text and the initial back projection position of the number; map the endpoint coordinates of the line segments corresponding to the arrow segments and the leader line segments to the original image content layer to generate the initial endpoint positions of the arrow segments and the leader line segments; map the rectangle boundary coordinates corresponding to the frame line segments and the curve boundary coordinates corresponding to the selection line to the original image content layer to generate the initial coverage area of ​​the frame line and the initial coverage area of ​​the selection line. Object boundary response extraction is performed on the original image content layer to generate an object boundary response map. Specifically, the original image content layer is subjected to grayscale conversion, and the grayscale change value of each pixel in the horizontal and vertical directions is calculated. The sum of the absolute values ​​of the grayscale change values ​​in the horizontal and vertical directions is taken as the boundary response value of the corresponding pixel. The average boundary response value of all pixels in the original image content layer is calculated, and pixels with boundary response values ​​greater than or equal to the average boundary response value are identified as object boundary pixels. Adjacent and connected object boundary pixels are aggregated to generate the object boundary response map. Based on the object boundary response map, the initial endpoint positions of the arrow segment and the initial endpoint positions of the leader segment are corrected to obtain the back projection positions of the arrow segment endpoints and the back projection positions of the leader segment endpoints. Endpoint repositioning correction includes: establishing endpoint search regions centered on the initial endpoint positions of the arrow segment and the leader segment, respectively; calculating the pixel distance between each object boundary pixel within the endpoint search region and its corresponding initial endpoint position; and reading the boundary response value corresponding to each object boundary pixel; calculating the endpoint repositioning value for each object boundary pixel within the endpoint search region, where the endpoint repositioning value is the boundary response value corresponding to the object boundary pixel divided by the sum of the pixel distance from the object boundary pixel to its corresponding initial endpoint position and one; determining the object boundary pixel with the largest endpoint repositioning value as the target repositioning pixel; and correcting the corresponding initial endpoint position to the target repositioning pixel position to obtain the arrow segment endpoint back projection position and the leader segment endpoint back projection position; the endpoint search region is a pixel region centered on the corresponding initial endpoint position and with a radius of one-hundredth of the short side pixel length of the image to be identified; Based on the object boundary response map, boundary snapping correction is performed on the initial coverage area of ​​the frame line and the initial coverage area of ​​the delineation line to obtain the frame line coverage area and the delineation coverage area. Boundary snapping correction includes: extracting the boundary pixels corresponding to the initial coverage areas of the bounding box and the initial coverage areas of the selection line; establishing boundary search areas centered on the boundary pixels of each region; reading the boundary response values ​​corresponding to the boundary pixels of each object within the boundary search area; filtering the boundary pixels of the object within the boundary search area whose boundary response values ​​are greater than or equal to the snapping response threshold; calculating the boundary snapping value of each filtered object boundary pixel, whereby the boundary snapping value is the boundary response value of the corresponding object boundary pixel divided by the sum of the pixel distance from the object boundary pixel to the corresponding region boundary pixel and one; determining the object boundary pixel with the largest boundary snapping value as the target snapping pixel, and replacing the corresponding region boundary pixel position with the target snapping pixel; and obtaining the bounding box coverage area and the selection coverage area after completing the position correction for all region boundary pixels; the snapping response threshold is the average value of the boundary response values ​​corresponding to all object boundary pixels in the object boundary response map; and the boundary search area is a pixel area centered on the corresponding region boundary pixel with a radius of one-hundredth of the short side pixel length of the image to be recognized. Based on the initial back projection position of the annotation text, the initial back projection position of the number, and the object boundary response map, the location correction of the neighboring object is performed to obtain the back projection position of the annotation text and the back projection position of the number. The back projection positions of the annotation text, the number, the back projection positions of the arrow segment endpoints, the back projection positions of the leader line segment endpoints, the frame coverage area, and the circled coverage area are encapsulated to generate the annotation element back projection result. The neighboring object localization correction includes: establishing object search regions centered on the initial back projection positions of the labeled text and the initial back projection positions of the numbers, respectively; extracting the set of object boundary pixels within the object search region; calculating the pixel distances between the initial back projection positions of the labeled text and each object boundary pixel, and the pixel distances between the initial back projection positions of the numbers and each object boundary pixel; calculating the neighboring location value of each object boundary pixel within the object search region, where the neighboring location value is the boundary response value corresponding to the object boundary pixel divided by the sum of the pixel distance from the object boundary pixel to the corresponding initial back projection position and one; determining the object boundary pixel with the largest neighboring location value as the target location pixel; and correcting the initial back projection positions of the labeled text and the numbers to the neighboring positions of the corresponding target location pixels, respectively, to obtain the labeled text back projection position and the number back projection position; the object search region is a pixel region centered on the corresponding initial back projection position with a radius of two percent of the short side pixel length of the image to be identified; the neighboring position of the target location pixel is the pixel position after offsetting by a preset offset distance from the target location pixel in the direction away from the object boundary region, where the preset offset distance is five-thousandths of the short side pixel length of the image to be identified.

[0025] In this embodiment, the generation of image annotation information recognition results specifically includes: The original image content layer is subjected to object region extraction to obtain an object region set, and the object region position and object region boundary are configured for each object region in the object region set. The object region extraction includes: performing region aggregation on the connected pixel regions in the original image content layer, and extracting connected regions whose region area is greater than the object area threshold as object regions, wherein the object area threshold is one-thousandth of the total number of pixels in the original image content layer. Based on the back projection results of the annotation elements and the object region set, candidate relationships of annotation objects are constructed, and the object association value corresponding to each candidate relationship of annotation objects is calculated. The candidate relationships of annotation objects include the candidate relationship between annotation text and object region, the candidate relationship between number and object region, the candidate relationship between arrow segment endpoint and object region, the candidate relationship between leader segment endpoint and object region, the candidate relationship between frame line coverage area and object region, and the candidate relationship between circle coverage area and object region. The object association value is calculated based on the relationship type corresponding to the candidate relationships of the labeled objects. The object association value corresponding to the candidate relationship between the endpoint of the arrow segment and the object area, and the candidate relationship between the endpoint of the leader segment and the object area, is the endpoint proximity value. The object association value corresponding to the candidate relationship between the bounding box area and the object area, and the candidate relationship between the circled area and the object area, is the overlap value. The object association value corresponding to the candidate relationship between the labeled text and the object area is the text proximity value. The object association value corresponding to the candidate relationship between the number and the object area is the number proximity value. The endpoint proximity value is the object boundary response value divided by the sum of the pixel distance from the endpoint of the segment to the boundary of the object area and one. The overlap value is the number of overlapping pixels between the bounding box area or the circled area and the object area divided by the number of pixels in the union. The text proximity value is the object boundary response value divided by the sum of the pixel distance from the back projection position of the labeled text to the boundary of the object area and one. The number proximity value is the object boundary response value divided by the sum of the pixel distance from the back projection position of the number to the boundary of the object area and one. For each annotation element in the annotation element set, the object region with the largest object association value is determined as the annotated object, and the annotation element identifier and the annotated object identifier are written into the same annotation object relationship item to obtain the generated correspondence relationship. The correspondence relationship includes the correspondence between annotation text and annotated object, the correspondence between number and annotated object, the correspondence between arrow segment and annotated object, the correspondence between leader segment and annotated object, the correspondence between the frame coverage area and annotated object, and the correspondence between the circle coverage area and annotated object. The image annotation information recognition result is generated by associating the set of annotation elements, the annotated objects, and their corresponding relationships. The image annotation information recognition result includes annotation content, annotation type, annotation layer position, annotated objects, the position of annotated objects, and the correspondence between annotation objects. Among them, annotation content includes the character content corresponding to the annotation text and number; annotation type includes text annotation, arrow annotation, leader line annotation, box selection annotation, circle selection annotation, and numbered annotation; annotation layer position includes the element position coordinates and element boundary coordinates of the annotation element in the annotation information candidate layer; the annotated object includes the object region identifier that establishes a corresponding relationship with the annotation element; the annotated object position includes the object region position and object region boundary corresponding to the annotated object; and the annotation object correspondence includes the association relationship between the annotation element identifier and the annotated object identifier.

[0026] Example 1: To verify the feasibility of this invention in practice, it was applied to a medical image annotation and recognition scenario. When organizing CT images and lesion analysis images, staff typically add arrows, circles, bounding boxes, and numbered text to mark lesion locations and key observation areas. Because some annotations are overlaid on the original image using a semi-transparent color overlay, and medical images themselves suffer from texture noise, compression distortion, and complex grayscale backgrounds, traditional recognition methods are prone to problems such as misalignment between annotated and lesion areas, arrow pointing errors, and incorrect number correspondences, leading to decreased efficiency in subsequent image archiving and case analysis.

[0027] When applying this invention, the medical image to be identified is first acquired, and the image is then subjected to size normalization, color space conversion, noise suppression, and image enhancement processing to generate basic visual data. Subsequently, based on the basic visual data, HSV color space features, Lab color space features, linewidth consistency features, edge closure morphology features, and transparent overlay trace features are extracted. After identifying the region containing the added annotation content through multi-source visual trace response, a set of overlaid annotation regions is generated. Next, the set of overlaid annotation regions is input into a SegFormer-based annotation layer segmentation model. Layer decoherence separation processing is performed on the main lesion content and the subsequently added annotation content in the medical image, generating the original image content layer and the annotation information candidate layer, respectively. The system further identifies text, numbers, arrow segments, and selection lines in the annotation information candidate layer, and remaps the identified annotation elements to the original image content layer through layer separation mapping relationships.

[0028] During the backprojection of labeled elements, this invention performs adaptive position correction on arrow endpoints, frame boundaries, and selected areas based on the object boundary response in the original image content layer. This ensures that arrow segments accurately point to the edge of the lesion and that the selected area maintains consistency with the actual lesion contour. Simultaneously, through object region extraction and object association value calculation, a correspondence is established between labeled elements and the labeled lesion areas, generating image annotation information recognition results. Continuous application statistics show that this invention maintains stable recognition even in complex backgrounds, semi-transparent annotations, and compression distortion scenarios. The accuracy of positional matching between labeled elements and lesion areas is significantly improved, reducing the number of manual reviews and increasing the efficiency of medical image annotation and processing.

[0029] Table 1. Performance Comparison of the Invention and Traditional Image Annotation Recognition Methods

[0030] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0031] In terms of the accuracy of labeled area recognition, this invention achieves 94.1%, which is 4.7 percentage points higher than the 89.4% of traditional recognition methods. This improvement is due to the fact that this invention constructs multi-source visual trace features through HSV color space features, Lab color space features, texture interruption features, and transparent overlay trace features. This enables joint analysis of the visual differences between the added annotation content and the main content of the image, thereby improving the ability to recognize labeled areas in complex backgrounds.

[0032] Regarding the accuracy of arrow endpoint positioning, this invention achieves 93.2%, compared to 87.8% for traditional recognition methods. Simultaneously, the misalignment rate is reduced from 8.3% to 3.9%. These results demonstrate that, during the backprojection process of annotation elements, this invention adaptively adjusts the positions of arrow segment endpoints and the area covered by the frame line through object boundary response extraction, endpoint repositioning correction, and boundary snapping correction. This reduces the positional offset problem caused by layer separation, thus improving the consistency between the arrow's pointing position and the actual object boundary.

[0033] In terms of the accuracy of object association, this invention achieves 92.4%, a significant improvement compared to the 85.9% of traditional recognition methods. The corresponding number of manual reviews is reduced from 17 per 100 images to 9 per 100 images. This result demonstrates that this invention, by constructing candidate relationships for labeled objects and calculating endpoint proximity, overlap, text proximity, and number proximity values ​​respectively, achieves accurate association between different labeling types and object regions, thus reducing the need for manual reviews due to errors in object association.

[0034] In terms of accuracy in recognizing semi-transparent annotations and complex backgrounds, this invention achieves 89.3% and 90.6% accuracy respectively, both higher than traditional recognition methods. Meanwhile, the average processing time of this invention is 391 milliseconds per image, lower than the 428 milliseconds per image of traditional recognition methods. This is because this invention utilizes decoherence separation processing of annotation layers, dividing the added annotation content and the main image content into layers. This reduces the interference of complex background textures and compressed noise on the recognition process, improving recognition accuracy while reducing subsequent error correction processing time.

[0035] 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 recognizing image annotation information based on visual intelligence, characterized in that, Includes the following steps: Acquire and preprocess the image to be recognized to generate basic visual data for the image; Extracting multi-source visual trace features based on image-based visual data; Based on the multi-source visual trace features, identify the candidate pixel regions with superimposed annotations in the image to be identified, and generate a set of superimposed annotation regions with annotation type and region attributes; The set of superimposed labeled regions is input into the label layer segmentation model based on SegFormer. The label layer decoherence separation process is performed on the image to be recognized to generate the original image content layer and the label information candidate layer and record the layer separation mapping relationship. The candidate layer of annotation information is subjected to text, line segment and contour recognition to generate a set of annotation elements; According to the layer separation mapping relationship, the set of labeled elements is back-projected to the original image content layer, and adaptive correction of the label element back-projection is performed based on the object boundary response in the original image content layer to generate the label element back-projection result; Based on the back projection results of the labeled elements and the object regions in the original image content layer, the correspondence between the set of labeled elements and the labeled objects is determined, and the image labeling information recognition results are generated.

2. The image annotation information recognition method based on visual intelligence according to claim 1, characterized in that, The preprocessing includes size normalization, color space conversion, noise suppression, and image enhancement.

3. The image annotation information recognition method based on visual intelligence according to claim 1, characterized in that, The extraction of multi-source visual trace features specifically includes: HSV color space data and Lab color space data are extracted from basic visual data of images. HSV color space features are generated based on the hue distribution, saturation distribution and lightness distribution in the HSV color space data. Lab color space features are generated based on the lightness component distribution and chromaticity component distribution in the Lab color space data. Color change detection is performed on adjacent pixel regions in the basic visual data of the image. Based on the color change amplitude of adjacent pixel regions in the HSV color space and Lab color space, color channel abrupt change features are generated. Laplacian edge detection is performed on the basic visual data of the image to extract the line width distribution, continuous line length and line direction variation in the edge response region. Line width consistency features are generated based on the consistent state of the line width distribution, and edge closure morphology features are generated based on the closed contour distribution in the edge response region. We perform texture continuity analysis on local texture regions in basic visual data of images, and generate local texture interruption features based on texture direction interruption locations, texture gray-level abrupt change locations, and texture repetition pattern disruption locations. Based on the local brightness superposition state, background transparency state, and JPEG compression block residual distribution in the basic visual data of the image, transparent superposition trace features and JPEG compression residual difference features are generated. The transparent superposition trace features, JPEG compression residual difference features, HSV color space features, Lab color space features, color channel abrupt change features, line width consistency features, edge closure shape features, and local texture interruption features are combined into multi-source visual trace features.

4. The image annotation information recognition method based on visual intelligence according to claim 1, characterized in that, The generation of the set of labeled overlay regions specifically includes: The multi-source visual trace features are subjected to coordinate alignment and scale unification processing, and the processed multi-source visual trace features are mapped to the corresponding pixel regions of the image to be identified to generate a multi-source visual trace response map. Calculate the confidence score of the superimposed annotations for each pixel region in the image to be identified based on the multi-source visual trace response map; Candidate pixel regions for annotation overlay are selected based on the annotation overlay confidence score and the annotation overlay judgment threshold. Connectivity aggregation and boundary correction are then performed on the candidate pixel regions for annotation overlay to generate candidate annotation regions. Based on the text response, line segment direction, endpoint shape, closed boundary, number character response, and color distribution characteristics of the candidate annotation area, the candidate annotation area is identified to obtain the annotation type area. The annotation type includes text annotation, arrow annotation, leader line annotation, box selection annotation, circle selection annotation, numbered annotation, and color label. Perform region deduplication, region merging, and region attribute configuration on candidate annotation regions corresponding to different annotation types to generate a set of annotation overlay regions.

5. The image annotation information recognition method based on visual intelligence according to claim 1, characterized in that, The records that record the layer separation mapping relationship specifically include: A mask for the superimposed annotation regions is generated based on the set of superimposed annotation regions. The mask for the superimposed annotation regions is then fused with the basic visual data of the image to generate input data for annotation layer segmentation. The annotation layer segmentation input data is input into the SegFormer-based annotation layer segmentation model. The hierarchical Transformer encoder performs multi-stage feature extraction on the annotation layer segmentation input data to generate multi-scale visual features at different resolutions. Multi-scale visual features of different resolutions are input into a multilayer perceptron decoder for pixel-level layer classification, generating layer segmentation results. Based on the layer segmentation results, the pixels in the image to be identified are classified into layers to generate the original image content layer and the annotation information candidate layer. Boundary consistency correction is performed on the original image content layer and the annotation information candidate layer, and the pixel coordinate correspondence between the original image content layer and the annotation information candidate layer is recorded to generate the layer separation mapping relationship.

6. The image annotation information recognition method based on visual intelligence according to claim 1, characterized in that, The generation of the labeled element set specifically includes: Text region detection is performed on the candidate layer of annotation information. Character segmentation, character contour feature extraction and character recognition are performed on the detected text regions based on OCR to obtain the annotation text and number in the candidate layer of annotation information and obtain the corresponding text box position and text box boundary. Edge extraction and binarization are performed on the candidate layer of annotation information to generate a linear annotation edge map. Candidate annotation line segments are detected from the linear annotation edge map based on probabilistic Hough transform, and the endpoints, lengths and directions of the candidate annotation line segments are extracted. Perform endpoint structure analysis and connection relationship analysis on candidate annotation line segments. Configure line segments with arrow endpoint structures as arrow line segments, configure line segments connecting the annotation text and the direction of the main content of the image as leader line segments, and configure line segments that connect with each other and form a rectangular boundary as frame line segments. Contour detection is performed on the candidate layer of annotation information to extract closed contours and color-connected contours. Contours with curved enclosing shapes in closed contours are configured as selection lines, and contours with continuous color distribution and regional filling features are configured as color markers. The annotation text, number, arrow segments, leader segments, frame segments, selection lines, and color identifiers are merged to generate a set of annotation elements. Based on the text box position, text box boundary, line segment endpoint coordinates, contour boundary coordinates, and identification category of each annotation element, the element position, element boundary, and element type are configured for each annotation element in the annotation element set.

7. The image annotation information recognition method based on visual intelligence according to claim 1, characterized in that, The generation of the back projection results of the labeled elements specifically includes: Read the layer separation mapping relationship, map the element positions and element boundaries in the annotation element set from the annotation information candidate layer to the original image content layer, and generate the initial back projection position of the annotation text, the initial back projection position of the number, the initial endpoint position of the arrow segment, the initial endpoint position of the leader segment, the initial coverage area of ​​the frame line, and the initial coverage area of ​​the circle line. Object boundary response is extracted from the content layer of the original image to generate an object boundary response map; Based on the object boundary response map, the initial endpoint positions of the arrow segment and the initial endpoint positions of the leader segment are corrected to obtain the back projection positions of the arrow segment endpoints and the back projection positions of the leader segment endpoints. Based on the object boundary response map, boundary snapping correction is performed on the initial coverage area of ​​the frame line and the initial coverage area of ​​the delineation line to obtain the frame line coverage area and the delineation coverage area. Based on the initial back projection position of the annotation text, the initial back projection position of the number, and the object boundary response map, the location correction of the neighboring object is performed to obtain the back projection position of the annotation text and the back projection position of the number. The results of the back projection positions of the annotation text, the number, the arrow segment endpoints, the leader segment endpoints, the frame coverage area, and the circled coverage area are then encapsulated to generate the annotation element back projection result.

8. The image annotation information recognition method based on visual intelligence according to claim 1, characterized in that, The generation of the image annotation information recognition result specifically includes: The object region is extracted from the content layer of the original image to obtain a set of object regions, and the object region position and object region boundary are configured for each object region in the set of object regions. Based on the back projection results of the labeled elements and the set of object regions, candidate relationships of labeled objects are constructed, and the object association value corresponding to each candidate relationship of labeled objects is calculated; For each annotation element in the annotation element set, the object region with the largest object association value is determined as the annotated object, and the annotation element identifier and the annotated object identifier are written into the same annotation object relationship item to obtain the generated correspondence; The image annotation information recognition result is generated by associating the annotation element set, the annotated object, and the corresponding relationship. The image annotation information recognition result includes the annotation content, annotation type, annotation layer position, annotated object, the position of the annotated object, and the correspondence between the annotated objects.

Images