A business audit-oriented image analysis result consistency control method

Abstract

The application discloses a kind of image analysis result consistency control methods for business audit, specifically related to edge image analysis and business audit access control technical field, including obtaining first graph that edge end is collected for auditing matter, first graph is conclusion generation graph;Split first graph, determine object area, reference area and doubt area, output carry conclusion type pending effective conclusion and first data;First data includes object area coordinate, reference area coordinate, doubt area coordinate, target element type, doubt area element type and target element mark;Reference area is the stable structure area with doubt area in object area and establishes doubt area relative coordinate;Entity element is attached with object area, non-entity element is not attached with object area.By pending effective conclusion trigger counterproof shooting action, and with entity / non-entity comparison field control conclusion access, when key counterproof factor is not excluded, no output is determined audit conclusion, relieve edge end premature conclusion problem.

Description

Technical Field

[0001] This invention relates to the field of edge image analysis and business review access control technology, and more specifically, to a method for controlling the consistency of image analysis results for business review. Background Technology

[0002] In image processing systems for business review, existing solutions often use edge devices as front-end judgment nodes. They analyze documents, vehicle damage, packaging damage, or on-site violations in the collected images through target detection, region recognition, image quality evaluation, and confidence level determination. When the recognition results meet preset conditions, they generate review prompts such as pass, abnormal, retake, or re-review. Taking on-site inspection of insurance claims as an example, inspectors usually use mobile terminals or vehicle edge devices to take pictures of the accident vehicle. The system needs to quickly determine whether there is any claimable damage to the vehicle body in a weak network environment. It is also limited by the time spent at the scene, upload bandwidth, user waiting time and privacy compliance requirements. It is impossible to upload every original image to the central end for complete recalculation, and it cannot rely on manual item-by-item verification in the long term. Under these conditions, when the review conclusion is output based solely on the detection frame, damage category, confidence level, or overall image clarity, the same local bright spots, paint reflections, stains, shadows, or texture changes caused by the shooting angle may be repeatedly judged as dents, scratches, or damage. Moreover, such misjudgments are often not low-confidence results; the edge end will write them as confirmed damage into the review process, only to overturn them after subsequent reshoots, reviews, or collection from different angles, resulting in inconsistent review conclusions. The reason is that the existing edge image analysis process mainly verifies whether positive visual features exist, but does not confirm whether the key counter-evidence factors sufficient to refute the conclusion have been excluded before the conclusion takes effect; The technical problem this application aims to solve is: how to implement access control for the process of converting image analysis results into business review conclusions under edge computing conditions, so that the edge end does not output a definitive review conclusion when key counter-evidence factors have not been eliminated. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for controlling the consistency of image analysis results for business review. This method involves recording the initial review image collected at the edge for the review item as the first image, and generating a pending conclusion, reference area, suspicious area, and target element based on the first image. A second image collection requirement is generated according to the pending conclusion and target element, and the counter-evidence verification image collected according to the second image collection requirement is recorded as the second image. Then, the method controls whether the pending conclusion is output as a confirmed review conclusion based on the registration results of the reference area between the first and second images, the mutual information sequence, and the coordinate difference sequence, thereby solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for controlling the consistency of image analysis results for business review, comprising: S1. Obtain the first image collected by the edge for the review items. The first image is the conclusion generation image. Segment the first image to determine the object area, reference area, and doubt area. Output the pending conclusion and first data carrying the conclusion type. The first data includes the coordinates of the object area, the coordinates of the reference area, the coordinates of the doubt area, the target element type, the doubt area element type, and the target element label. The pending conclusion is the judgment result generated based on the first image and not output as the final review conclusion. The doubt area is the local area that supports the pending conclusion. The reference area is a stable structural area located in the same object area as the doubt area and establishing the relative coordinates of the doubt area. S2. Read the counter-evidence shooting action according to the conclusion type, target element type and doubtful area element type, write the doubtful area coordinates, reference area coordinates and counter-evidence shooting action into the acquisition command, and output the second image acquisition requirements; the counter-evidence shooting action changes the acquisition angle, incident light direction, acquisition distance or occlusion relationship, so that non-physical elements change the imaging position or display state, and the physical elements maintain the reference area coordinates. S3. Obtain the second image, which is a counter-evidence verification image collected according to the requirements of the second image collection; locate the second reference area, establish a registration relationship with the first reference point column and the second reference point column, and map the coordinates of the suspected area to the second image to obtain the verification area; S4. According to the suspicious area element type and target element label, extract entity elements and non-entity elements in the suspicious area and verification area to form entity coordinate sequence and non-entity coordinate sequence, and obtain coordinate comparison results. S5. When the entity element forms a mutual information sequence and a coordinate difference sequence in the coordinate comparison result, the missing sequence number is empty, and there is no non-entity mutual information sequence and non-entity coordinate difference sequence, the pending conclusion is the confirmed review conclusion. When the mutual information sequence, coordinate difference sequence or missing sequence number of the entity element does not meet the output conditions, or when there is a non-entity mutual information sequence or non-entity coordinate difference sequence but no entity mutual information sequence and entity coordinate difference sequence, the confirmed review conclusion is stopped, and a supplementary collection instruction or a partial recalculation instruction is output.

[0005] In a preferred embodiment, S1 includes: S11. Obtain the first image collected by the edge end for the audit item, read the object category corresponding to the audit item, divide the outer contour of the audited object in the first image according to the object category, record the area enclosed by the outer contour as the object area, and extract the structural area with closed boundary, intersection point or continuous texture in the object area as the reference area. S12. Extract the local area that participates in generating the conclusion to be effective within the object area, and record the local area as the suspicious area. Extract the boundary line, texture line, bright spot area or occlusion area within the suspicious area. Record the boundary line, texture line, bright spot area or occlusion area that participates in generating the conclusion to be effective as the target feature. Output the coordinates of the suspicious area, the type of the target feature and the label of the target feature. S13. Using the closed boundary, intersection points, or continuous texture of the reference area as the coordinate reference, write the coordinates of the object area, the reference area, and the suspicious area into the same coordinate system; when the target feature type is a boundary line or texture line, and there is a connection relationship, relative position relationship, arrangement order, or connection direction between the target feature and the outer contour of the object area or the reference area point column, mark the target feature as a solid feature; when the target feature type is a bright spot area or an occluded area, mark the target feature as a non-solid feature, and output the conclusion to be effective and the first data.

[0006] In a preferred embodiment, S2 includes: S21. Read the conclusion type of the conclusion to be effective and the coordinates of the suspicious area, the coordinates of the reference area, the target element type, the suspicious area element type and the target element marker in the first data. Write the conclusion type into the conclusion field, write the suspicious area element type into the element field, write the target element marker into the target field, write the target element type into the type field, and generate the counter-evidence items according to the conclusion field, element field, target field and type field. S22. Determine the counter-evidence shooting action based on the counter-evidence item: When the type field is bright spot area, write the change of incident light direction into the counter-evidence shooting action; when the type field is boundary line or texture line, write the change of acquisition view and acquisition distance into the counter-evidence shooting action; when the type field is occlusion area, switch to the occlusion relationship change action generation process. S23. When the type field is bright spot area, boundary line or texture line, the acquisition boundary is formed by the coordinates of the suspicious area and the coordinates of the reference area. The acquisition boundary is written into the framing field, the counter-evidence shooting action is written into the action field, the coordinates of the reference area in the first data are written into the alignment field, the coordinates of the suspicious area are written into the verification field, the acquisition command is generated, and the acquisition command is output as the acquisition requirements of the second image.

[0007] In a preferred embodiment, S2 further includes: S24. Read the coordinates of the suspected area, the coordinates of the reference area in the first data, and the target element marker. Read the occlusion area segmentation result corresponding to the target element marker within the coordinates of the suspected area, and obtain the contour points of the occlusion area through grayscale edge detection, segmentation mask boundary extraction, or connected region outer contour extraction. Generate the contour center coordinates based on the horizontal and vertical coordinates of the contour points of the occlusion area. Record the direction opposite to the direction of the reference origin contained in the reference area coordinates in the first data as the occlusion direction. Write the occlusion direction into the occlusion field. Generate an occlusion relationship change action based on the occlusion field. Write the occlusion relationship change action into the action field. Write the coordinates of the suspected area into the verification field. Write the coordinates of the reference area in the first data into the alignment field. Generate an occlusion counter-evidence acquisition instruction. Output the occlusion counter-evidence acquisition instruction as the second image acquisition requirement.

[0008] In a preferred embodiment, S3 includes: S31. Obtain the second image collected according to the second image collection requirements; read the object area coordinates, reference area coordinates, and suspicious area coordinates from the first data; and read the counter-evidence shooting action from the second image collection requirements. In the second image, extract closed boundaries, intersection points, or continuous textures according to the reference type corresponding to the reference area of ​​the first image, and generate a second candidate point column according to the connection, adjacency, or extension relationship. In the reference area of ​​the first image, generate a first reference point column according to the same reference type, and mark the coordinates of the area where the first reference point column is located as the coordinates of the first reference area. The point column order is the sequence number relationship formed by the connection, adjacency, or extension relationship of each point in the point column. Record the point column in the second candidate point column that has the same point column order as the first reference point column as the second reference point column, mark the area where the second reference point column is located as the second reference area, and mark the coordinates of the area where the second reference point column is located as the coordinates of the second reference area. S32. Establish point pairs according to the points in the first reference point column and the second reference point column with the same order. Write the coordinate difference of each point pair into the registration displacement table. Based on the registration displacement table, map the object area coordinates and doubt area coordinates in the first data point by point to the second figure to generate the object area mapping area and the doubt area mapping area.

[0009] In a preferred embodiment, S3 further includes: S33. The overlapping area between the object area mapping region and the same audited object area in the second figure is recorded as the object confirmation area. The overlapping area between the suspicious area mapping region and the object confirmation region is recorded as the initial verification area. The retained objects in the initial verification area are adjusted according to the counter-evidence shooting action: when the incident light direction changes, the corresponding reference area point sequence is retained but the corresponding bright spot area coordinates are not retained; when the acquisition angle or acquisition distance changes, the corresponding reference area point sequence is retained but the corresponding suspicious area scale is not retained; when the occlusion relationship changes, the corresponding reference area point sequence is retained but the corresponding occlusion contour is not retained. The adjusted initial verification area is recorded as the verification area, and the verification area coordinates and registration displacement table are output.

[0010] In a preferred embodiment, S4 includes: S41. Read the coordinates of the suspicious area, the coordinates of the verification area, the coordinates of the reference area and the second reference area in the first data, the suspicious area feature type, the target feature type, the target feature marker, and the registration displacement table. Extract the first target feature point column in the suspicious area according to the target feature marker, and extract the second target feature point column in the verification area. When the target feature type is a boundary line, the first target feature point column and the second target feature point column are boundary points arranged in the order of line segment connection. When the target feature type is a texture line, the first target feature point column and the second target feature point column are texture points arranged in the order of texture extension. When the target feature type is a bright spot area, the first target feature point column and the second target feature point column are contour points arranged in the order of contour connection. S42. Using the coordinates of the reference area in the first data as a reference, calculate the coordinates of each point in the first target element point list relative to the coordinates of the reference area in the first data, and generate a first coordinate sequence; using the coordinates of the second reference area as a reference, calculate the coordinates of each point in the second target element point list relative to the coordinates of the second reference area, and generate a second coordinate sequence. S43. According to the target element marking and point sequence, divide the consecutive coordinate points with the same sequence number in the first coordinate sequence and the second coordinate sequence into corresponding segments. Read the gray value, edge direction and texture direction of the pixels where the coordinate points are located in the corresponding segments of the first coordinate sequence and the corresponding segments of the second coordinate sequence respectively. Count the occurrence of gray value, edge direction and texture direction. Divide the occurrence by the total number of pixels in the corresponding segment to obtain the probability value. Then sum the products of each probability value and the logarithm of each probability value after taking the opposite number to obtain the first entropy sequence and the second entropy sequence.

[0011] In a preferred embodiment, S4 further includes: S44. Take the gray value, edge direction, and texture direction of the pixels with the same index in the corresponding segment of the first coordinate sequence and the second coordinate sequence to form a joint value record. Count the number of occurrences of the joint value record. Divide the number of occurrences of the joint value record by the total number of joint value records in the corresponding segment to obtain the joint probability value. Take the negative of the product of each joint probability value and the logarithm of each joint probability value and sum them to obtain the joint entropy sequence. Add the first entropy sequence and the second entropy sequence segment by segment and then subtract the joint entropy sequence to obtain the mutual information sequence. S45. Map the first map coordinates of the first target feature point list to the verification area according to the registration displacement table, and convert them to the second reference area coordinates to obtain the reconstructed coordinate sequence. Subtract the coordinate points with the same sequence number in the reconstructed coordinate sequence from the coordinate points with the same sequence number in the second coordinate sequence to obtain the coordinate difference sequence. If the coordinate points with the same sequence number are missing, write the missing sequence number into the coordinate difference sequence. If the suspected feature type is an entity feature, write the mutual information sequence and the coordinate difference sequence into the entity comparison field. If the suspected feature type is a non-entity feature, write the mutual information sequence and the coordinate difference sequence into the non-entity comparison field. Merge the entity comparison field and the non-entity comparison field to generate the coordinate comparison result.

[0012] In a preferred embodiment, S5 includes: S51. Read the pending conclusion, the coordinates of the suspicious area, the target element marker and the coordinate comparison result. Read the entity comparison field and the non-entity comparison field from the coordinate comparison result. Extract the entity mutual information sequence, entity coordinate difference sequence and entity missing number corresponding to the target element marker from the entity comparison field. Extract the non-entity mutual information sequence, non-entity coordinate difference sequence and non-entity missing number corresponding to the target element marker from the non-entity comparison field. S52. When both the entity mutual information sequence and the entity coordinate difference sequence exist, the entity missing sequence number is empty, and neither the non-entity mutual information sequence nor the non-entity coordinate difference sequence exists, the pending conclusion is written into the confirmed audit conclusion and output.

[0013] In a preferred embodiment, S5 further includes: S53. When any of the following conditions are met: the entity mutual information sequence does not exist, the entity coordinate difference sequence does not exist, or the entity missing sequence number is not empty, stop outputting the confirmed audit conclusion; when a non-entity mutual information sequence or a non-entity coordinate difference sequence exists, and neither the entity mutual information sequence nor the entity coordinate difference sequence exists, stop outputting the confirmed audit conclusion; when the entity missing sequence number is not empty, write the suspicious area coordinates and the entity missing sequence number into the supplementary collection instruction; when the entity mutual information sequence does not exist and the entity coordinate difference sequence exists, or when the entity mutual information sequence exists and the entity coordinate difference sequence does not exist, write the target element mark and the suspicious area coordinates into the local recalculation instruction.

[0014] The technical effects and advantages of this invention are as follows: By triggering the action of taking counter-evidence photos through the pending conclusion, and controlling the admission of the conclusion with the entity / non-entity comparison field, a definitive review conclusion is not output when key counter-evidence factors are not eliminated, thus alleviating the problem of premature conclusions at the edge. By establishing the coordinate correspondence between the two images through the reference area, the second reference area and the registration displacement table, the suspicious area is mapped to the verification area, so that the subsequent comparison is limited to the same object range, reducing the interference of the background or misaligned area. By separating the records of target feature types and suspicious feature types, the boundary lines, texture lines, bright spots and occluded areas are imported into the corresponding processing paths, making the verification logic of entity features and non-entity features clearer. By calculating the information correspondence of target elements using entropy, joint entropy, and mutual information sequences, and combining this with the reconstructed coordinate difference sequence, the image content association and spatial correspondence status can be reflected simultaneously. The missing sequence number field records situations where the target element does not form a corresponding point or segment, and generates supplementary collection instructions accordingly, so that the supplementary collection range is relatively concentrated in the suspicious area and the missing position. Distinguishing between entity fields, non-entity fields, and incomplete field states helps to determine output, intercept anomalies, and perform partial recalculation, which helps reduce the reliance on recalculating the entire drawing or manual review. Attached Figure Description

[0015] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Refer to the instruction manual appendix Figure 1 The present invention provides a method for controlling the consistency of image analysis results for business review, comprising: S1. Obtain the first image collected by the edge for the review items. The first image is the conclusion generation image. Segment the first image to determine the object area, reference area, and doubt area. Output the pending conclusion and first data carrying the conclusion type. The first data includes the coordinates of the object area, the coordinates of the reference area, the coordinates of the doubt area, the target element type, the doubt area element type, and the target element label. The pending conclusion is the judgment result generated based on the first image and not output as the final review conclusion. The doubt area is the local area that supports the pending conclusion. The reference area is a stable structural area located in the same object area as the doubt area and establishing the relative coordinates of the doubt area. This implementation method is used to generate the first data required for subsequent counter-evidence collection, registration and positioning, and coordinate comparison in the first image stage. The first image is the initial review image collected at the edge for the review matters and used to generate the conclusion to be effective. The image analysis results of the first image include the conclusion to be effective, the local area involved in generating the conclusion to be effective, and candidate information of target elements. This implementation method first determines the object area based on the review matters, then selects a reference area within the object area, avoiding the local area involved in generating the conclusion to be effective. Subsequently, target elements are identified in the suspicious area, and the target element type, target element marker, and suspicious area element type are written into the first data. Among them, the target element type records the boundary line, texture line, bright spot area, or occluded area; the target element marker records the specific element involved in generating the conclusion to be effective; and the suspicious area element type records the initial entity / non-entity attribution corresponding to the target element marker, which is not used as the final verification conclusion. This implementation process includes the following steps: S11 is used to determine the object area and reference area, providing a coordinate basis for establishing the relative coordinates of the suspected area and subsequent alignment and positioning. After the edge end acquires the first image collected for the audit item, it reads the object category corresponding to the audit item, segments the outer contour of the audited object in the first image according to the object category, records the area enclosed by the outer contour as the object area, and writes the object area coordinates into the object area coordinate field of the first data. The edge end then reads the local area in the image analysis result of the first image that participates in generating the conclusion to be effective. After excluding the local area in the object area, it extracts the structural area with closed boundaries, intersection points, or continuous textures. The structural area is recorded as the reference area, and the reference area coordinates are written into the reference area coordinate field of the first data. The reference area is not composed of the local area or specific elements that participate in generating the conclusion to be effective. The coordinates of the reference area are used as the alignment basis when generating subsequent collection requirements. When closed boundaries, intersection points, or continuous textures are not extracted in the object area, the edge end does not write the reference area coordinates, does not generate subsequent collection requirements, and outputs a prompt to re-collect the first image. S12 is used to determine the local basis for forming the conclusion to be effective from the object area, and to transform the local basis into target elements that can be tracked by the marker in subsequent coordinate comparison; the edge end reads the local area involved in generating the conclusion to be effective in the object area, records the local area as the suspicious area, and writes the coordinates of the suspicious area into the suspicious area coordinate field of the first data; then, it identifies the boundary line, texture line, bright spot area or occlusion area in the suspicious area, writes the specific elements involved in generating the conclusion to be effective into the target element marker, and writes the category of the specific element into the target element type; the target element type records the boundary line, texture line, bright spot area or occlusion area, and the target element marker records the specific elements involved in generating the conclusion to be effective; when the target element type is a boundary line, the edge end... Boundary points are extracted in the order of line segment connection. When the target feature type is a texture line, texture points are extracted at the edge end in the order of texture extension. When the target feature type is a bright spot area, contour points are extracted at the edge end in the order of contour connection. When the target feature type is an occluded area, local occluded contour points covering the boundary line, texture line, or bright spot area are extracted at the edge end. The occluded area is used as a control element for reverse evidence acquisition. When performing coordinate comparison after the occlusion relationship changes, the boundary line, texture line, or bright spot area exposed after the occlusion is removed is processed, and the occluded area itself is not used as the basis for confirming the entity feature. If no target feature is extracted in the suspected area to participate in generating the conclusion to be effective, the edge end records the target feature extraction anomaly and re-executes the target feature extraction in the suspected area or supplements the target feature extraction data. S13 is used to write the object area, reference area, and suspected area into the same coordinate system and form the initial entity / non-entity attribution of the target element; the edge end establishes a coordinate reference based on the reference area: when the reference area is a closed boundary, the starting point of the closed boundary point sequence is taken as the reference origin, and the connection direction from the starting point to the adjacent boundary point is taken as the reference direction; when the reference area is an intersection point, the intersection point is taken as the reference origin, and the connection direction from the intersection point to the adjacent structural point is taken as the reference direction; when the reference area is a continuous texture, the starting point of the continuous texture point sequence is taken as the reference origin, and the texture extension direction is taken as the reference direction; the edge end converts the object area coordinates, reference area coordinates, and suspected area coordinates to the same coordinate system based on the reference origin and reference direction, and writes the first data; the edge end then reads the target element marker and target element type; when the target element is... When the target feature type is a boundary line or texture line, and the target feature point series has a connection relationship with the outer contour of the object area and the reference area point series, or when the position, arrangement order, or connection direction of the target feature point series relative to the coordinates of the reference area corresponds to scratches, cracks, damaged edges, or texture breaks, the target feature mark is written into the initial assignment of the entity feature; when the target feature type is a bright spot area or an occluded area, the target feature mark is written into the initial assignment of the non-entity feature; the suspicious area feature type records the entity / non-entity initial assignment corresponding to the target feature mark; S13 outputs the pending conclusion and the first data, which includes the coordinates of the object area, the coordinates of the reference area, the coordinates of the suspicious area, the target feature type, the suspicious area feature type, and the target feature mark, for subsequent generation of acquisition requirements, location registration area, and execution of coordinate comparison; In this implementation method, the first data output by S1 forms the field entry for subsequent steps: the reference area coordinates serve as the alignment basis for subsequent acquisition requirements and registration positioning; the suspicious area coordinates serve as the verification basis for subsequent acquisition requirements and suspicious area mapping; the target element type is used to determine the counter-evidence shooting action; the target element marker is used to extract the corresponding target element point sequence; the suspicious area element type is used to determine the mutual information sequence, coordinate difference sequence, and missing sequence number to be written into the entity comparison field or non-entity comparison field; in practical applications: taking vehicle damage review as an example, after obtaining the first image of the car door at the edge, the outer contour of the car door is segmented according to the vehicle exterior parts category to obtain the object area, and the door frame is... Reference areas are extracted at intersections or continuous body lines, and the local area where the suspected damage is located is recorded as the suspected area. If a continuous scratch boundary line is extracted within the suspected area, the target element type is boundary line, the target element marker records the continuous scratch boundary line, and the suspected area element type is recorded as a solid element. Subsequently, actions to change the acquisition viewpoint or acquisition distance are generated based on the target element type and the suspected area element type. If a bright spot area formed by reflection is extracted within the suspected area, the target element type is bright spot area, the target element marker records the bright spot area, and the suspected area element type is recorded as a non-solid element. Subsequently, actions to change the incident light direction are generated based on the target element type and the suspected area element type.

[0018] S2. Read the counter-evidence shooting action according to the conclusion type, target element type and doubtful area element type, write the doubtful area coordinates, reference area coordinates and counter-evidence shooting action into the acquisition command, and output the second image acquisition requirements; the counter-evidence shooting action changes the acquisition angle, incident light direction, acquisition distance or occlusion relationship, so that non-physical elements change the imaging position or display state, and the physical elements maintain the reference area coordinates. This implementation method is used to generate second image acquisition requirements based on the conclusion to be effective and the first data, so that the acquisition end can obtain the counter-evidence verification image according to the second image acquisition requirements; the edge end first reads the conclusion type carried by the conclusion to be effective, and reads the coordinates of the suspicious area, the coordinates of the reference area, the target element type, the suspicious area element type, and the target element marker in the first data to generate counter-evidence items; then, the counter-evidence shooting action is determined based on the counter-evidence items; in the case of no obstruction, the framing range, action requirements, alignment basis, and verification area are written into the acquisition command; in the case of obstruction, the obstruction direction is generated according to the outline of the obstructed area and the coordinates of the reference area, and the obstructed counter-evidence acquisition command is generated; the implementation process includes the following steps: S21 is used to convert the fields in the pending conclusion and the first data into counter-evidence items, providing a field basis for the selection of counter-evidence shooting actions; the edge end reads the pending conclusion output by S1, reads the conclusion type from the pending conclusion, and reads the suspicious area coordinates, reference area coordinates, target element type, suspicious area element type, and target element marker from the first data; writes the conclusion type into the conclusion field, the suspicious area element type into the element field, the target element marker into the target field, and the target element type into the type field, and generates counter-evidence items according to the conclusion field, element field, target field, and type field; wherein, the conclusion field records the review judgment category to which the pending conclusion belongs, the type field records the boundary line, texture line, bright spot area, or occluded area, the element field records the initial affiliation of entity elements or non-entity elements, and the target field records the specific target elements that participated in generating the pending conclusion; when the suspicious area coordinates, reference area coordinates, target element type, suspicious area element type, or target element marker is missing, the edge end records the field missing status, does not generate counter-evidence items, and outputs a supplementary prompt for the first data; S22 is used to determine the counter-evidence shooting action based on the counter-evidence item, so that the target element type corresponding to the acquisition conditions of the second image changes; the edge end reads the type field and element field in the counter-evidence item; when the type field is a bright spot area and the element field is recorded as a non-solid element, the incident light direction change is written into the counter-evidence shooting action; when the type field is a boundary line or texture line and the element field is recorded as a solid element, the acquisition viewpoint change and acquisition distance change are written into the counter-evidence shooting action. The acquisition viewpoint change is used to change the imaging angle of the target element relative to the acquisition direction, and the acquisition distance change is used to change the imaging scale of the target element in the second image; when the type field is an occluded area, the edge end does not generate a normal counter-evidence shooting action, and enters the occlusion relationship change action generation process; when the type field does not record a boundary line, texture line, bright spot area or occluded area, the edge end records the abnormal state of the target element type and outputs a target element supplement prompt; S23 is used to write the counter-evidence shooting action, suspicious area coordinates, and reference area coordinates into the acquisition command in the case of no obstruction, forming the second image acquisition requirements; when the type field is bright spot area, boundary line, or texture line, the edge end reads the suspicious area coordinates and reference area coordinates, merges all boundary coordinate points in the suspicious area coordinates and all boundary coordinate points in the reference area coordinates into a set of acquisition points; reads the horizontal coordinate value endpoints and vertical coordinate value endpoints from the set of acquisition points, connects them according to the left endpoint, upper endpoint, right endpoint, and lower endpoint to form the acquisition boundary surrounding the suspicious area and reference area, and writes the acquisition boundary into the framing field; the edge end writes the counter-evidence shooting action into the action field, writes the reference area coordinates into the alignment field, writes the suspicious area coordinates into the verification field, generates the acquisition command, and outputs the acquisition command as the second image acquisition requirements; among them, the alignment field records the acquisition requirements of the area corresponding to the reference area, and the verification field records the acquisition requirements of the area corresponding to the suspicious area; when the acquisition point set lacks suspicious area boundary coordinate points or reference area boundary coordinate points, the edge end records the field missing status and outputs the first data supplement prompt; S24 is used to generate an occlusion counter-evidence acquisition command in occlusion situations, causing the occluded area to leave the suspected area in the occlusion direction, while retaining the corresponding area of ​​the reference area; when the type field is occlusion area, the edge end reads the coordinates of the suspected area, the coordinates of the reference area, and the target feature marker, reads the occlusion area segmentation result corresponding to the target feature marker within the suspected area coordinates, and obtains the occlusion area contour points through grayscale edge detection, segmentation mask boundary extraction, or connected region outer contour extraction; the edge end generates the contour center coordinates based on the horizontal and vertical coordinates of the occlusion area contour points, and reads the reference origin in the reference area coordinates; the direction of the contour center coordinates pointing to the reference origin is recorded as the contour pointing direction, and the opposite direction of the contour pointing direction is recorded as the occlusion direction, and the occlusion direction is written into the occlusion field; the edge end generates an occlusion relationship change action based on the occlusion field, changing the occlusion relationship... The action is written to the action field, the coordinates of the suspected area are written to the verification field, and the coordinates of the reference area are written to the alignment field. Based on the coordinates of the suspected area and the coordinates of the reference area, the framing field is generated according to the acquisition boundary generation method in S23 to obtain the occlusion counter-evidence acquisition instruction. The edge end outputs the occlusion counter-evidence acquisition instruction as the second image acquisition requirement, so that the acquisition end obtains a second image containing the corresponding area of ​​the reference area and the corresponding area of ​​the suspected area, and after the occlusion relationship changes, the occluded boundary line, texture line or bright spot area in the suspected area is revealed. The occlusion area contour is only used to determine the occlusion direction and the area to be moved out, and is not used as the basis for the confirmation of entity elements in the subsequent S4. When the occlusion area contour point is not extracted, the edge end records the occlusion contour missing status and outputs the occlusion contour supplementation prompt. When the reference area coordinates are missing, the edge end records the field missing status and outputs the first data supplementation prompt. Through this implementation method, S2 transforms the conclusion to be effective and the conclusion type, target element type, suspicious area element type, target element marker, suspicious area coordinates, and reference area coordinates in the first data into the second image acquisition requirements; the conclusion field, element field, target field, and type field are used to generate counter-evidence items; the framing field limits the acquisition range of the second image; the alignment field is used for subsequent positioning of the corresponding area of ​​the reference area; the verification field is used for subsequent determination of the corresponding area of ​​the suspicious area; the action field records changes in acquisition viewpoint, acquisition distance, incident light direction, or occlusion relationship; the occlusion field is used for directional control of occlusion branches; in practical applications: taking vehicle damage review as an example, if the target element type in the first data is a bright spot area or a suspicious area... If the feature type is a non-solid feature, the edge end will write the change of the incident light direction into the action field and generate a second image acquisition requirement covering the door waistline reference area and the suspected bright spot area; if the target feature type in the first data is a boundary line and the suspected area feature type is a solid feature, the edge end will write the change of acquisition viewpoint and acquisition distance into the action field and require the second image to retain the door waistline reference area and the suspected continuous scratch area; if the target feature type in the first data is an occluded area, the edge end will determine the occlusion direction based on the occlusion area outline point and the reference origin in the reference area coordinates, generate an occlusion counter-evidence acquisition instruction, so that the occluded area moves away from the suspected area along the occlusion direction, and the suspected damage location of the door and the door waistline reference area enter the framing range of the second image.

[0019] S3. Obtain the second image, which is a counter-evidence verification image collected according to the requirements of the second image collection; locate the second reference area, establish a registration relationship with the first reference point column and the second reference point column, and map the coordinates of the suspected area to the second image to obtain the verification area; This implementation method is used to locate the area corresponding to the reference area in the second image as the second reference area, and to map the coordinates of the suspicious area in the first data to the second image to form a verification area. The edge end first reads the second image acquisition requirements to obtain the second image, and extracts the second candidate point column in the second image according to the reference type corresponding to the reference area in the first image. Then, the second candidate point column is matched with the first reference point column in the order of the point column to determine the second reference area. Subsequently, a registration displacement table is generated based on the points with the same order in the first and second reference point columns, and the coordinates of the object area and the suspicious area are mapped to the second image. The verification area is obtained by combining the same object area in the second image. This implementation process includes the following steps: S31 is used to locate the second reference area in the second image, providing a second reference point list for subsequent generation of the registration displacement table; after the edge end acquires the second image acquired according to the second image acquisition requirements, it reads the object area coordinates, reference area coordinates, and suspicious area coordinates from the first data, and reads the framing field, alignment field, and reverse evidence shooting action from the second image acquisition requirements; within the area defined by the framing field of the second image, the edge end extracts closed boundaries, intersection points, or continuous textures according to the reference type corresponding to the reference area of ​​the first image, and generates a second candidate point list carrying the reference type according to the connection, adjacency, or extension relationship; a first reference point list is generated in the reference area of ​​the first image according to the same reference type, wherein the order of the point list is the ordinal relationship formed by the connection, adjacency, or extension relationship of each point in the point list; when the reference type is a closed boundary When the reference type is a corner point, the first reference point column and the second candidate point column are arranged in the order of boundary point connection; when the reference type is a corner point, the first reference point column and the second candidate point column are arranged in the order of adjacency between the corner point and the adjacent structural point; when the reference type is a continuous texture, the first reference point column and the second candidate point column are arranged in the order of texture extension; the edge end records the point column in the second candidate point column that has the same point column order as the first reference point column as the second reference point column, records the area where the second reference point column is located as the second reference area, and writes the coordinates of the second reference area; if the second candidate point column corresponding to the reference area of ​​the first image is not extracted in the second image, or if there is no point column in the second candidate point column that has the same point column order as the first reference point column, the edge end does not generate the coordinates of the second reference area and outputs a prompt for re-acquisition of the second image; S32 is used to generate a registration displacement table based on the first and second reference point columns, and map the object area coordinates and suspected area coordinates in the first data to the second map; the edge end establishes point pairs according to the points in the first and second reference point columns with the same order, reads the coordinates of the first and second maps for each pair, and subtracts the coordinates of the first map from the coordinates of the second map to obtain the point pair coordinate difference; the edge end writes the point pair number, the coordinates of the first map, the coordinates of the second map, and the point pair coordinate difference into the registration displacement table, which serves as the data form of the registration relationship; subsequently, the edge end reads the object area coordinates and suspected area coordinates in the first data, and for the points to be mapped in the object area coordinates and suspected area coordinates, reads the corresponding displacement item according to the position of the point to be mapped relative to the connecting segment of the first reference point column; the point to be mapped... When the point to be mapped is located at a reference point in the first reference point column, the point-to-point coordinate difference of the corresponding point pair is read. When the point to be mapped is located between two adjacent reference points in the first reference point column, the point-to-point coordinate difference of the reference points at both ends of the connecting segment is read, and the corresponding displacement term is read according to the index position of the point to be mapped in the connecting segment relative to the two end reference points. The edge end adds the corresponding displacement term to the coordinates of the point to be mapped to obtain the coordinates of the mapped point in the second figure. The object area coordinates are mapped point by point to generate the object area mapping region, and the suspicious area coordinates are mapped point by point to generate the suspicious area mapping region. If the position of the point to be mapped relative to the connecting segment of the first reference point column cannot be determined, the edge end writes the index of the point to be mapped into the mapping missing record. The mapping missing record is used for the missing index in the subsequent S4 generation of the coordinate difference sequence. S33 is used to limit the suspicious area mapping region to the same object range in the second image and form a verification area based on the counter-evidence shooting action; the edge end, according to the same object category as S11, divides the outer contour of the object under review in the second image, and records the area enclosed by the outer contour as the second object area; the edge end forms an overlapping area by intersecting the object area mapping region and the second object area according to the coordinate points, and records this overlapping area as the object confirmation area; then, the suspicious area mapping region and the object confirmation area form an overlapping area by intersecting the coordinate points, and records this overlapping area as the initial verification area; for suspicious area mapping points in the suspicious area mapping region that do not fall into the object confirmation area, the edge end deletes the suspicious area mapping point from the initial verification area and writes the sequence number of the deleted suspicious area mapping point into the mapping missing record; the edge end then adjusts the retained objects in the initial verification area according to the counter-evidence shooting action: when the incident light direction changes, retain The coordinate correspondence of the reference area is not retained, but the coordinate correspondence of the bright spot area is not retained. When the acquisition viewpoint or acquisition distance changes, the coordinate correspondence of the reference area point sequence is retained, but the coordinate correspondence of the suspicious area scale is not retained. When the occlusion relationship changes, the coordinate correspondence of the reference area point sequence is retained, but the coordinate correspondence of the occlusion contour is not retained. Among them, not retaining the coordinate correspondence of the bright spot area, the coordinate correspondence of the suspicious area scale, or the coordinate correspondence of the occlusion contour means that the object does not participate in determining the boundary of the second reference area, the registration displacement table, or the verification area, and does not mean that the object will not enter the subsequent S4 extraction. The edge end records the adjusted initial verification area as the verification area and outputs the coordinates of the verification area and the registration displacement table. If the second object area is not segmented in the second image, the edge end does not generate the object confirmation area and the initial verification area, and outputs the second image re-acquisition prompt. If the object confirmation area exists but the coordinate points of the suspicious area mapping area and the object confirmation area do not intersect, the edge end outputs the suspicious area mapping failure prompt and does not enter S4. In this implementation method, S3 determines the area corresponding to the reference area in the second image as the second reference area, and writes the registration relationship from the coordinates of the first image to the coordinates of the second image into the registration displacement table; the coordinates of the object area are mapped to the object area mapping area through the registration displacement table, and the coordinates of the suspected area are mapped to the suspected area mapping area through the registration displacement table. The suspected area mapping area intersects with the object confirmation area to form the verification area; the second reference area coordinates, verification area coordinates, and registration displacement table output by S3 are read by S4. Among them, the second reference area coordinates are used to generate the coordinate reference of the second coordinate sequence, and the registration displacement table is used to reconstruct the first target element point series to the verification area; in practical applications: taking vehicle damage review as an example, the second image acquisition requirement requires that the second image retain the door. The waistline reference area and the suspected damage area are defined by generating a second candidate point column carrying a continuous texture reference type according to the continuous texture extension order of the waistline of the car door within the framing field of the second image. This candidate point column is then matched with the first reference point column of the waistline of the car door in the first image according to the point column order to obtain the second reference area. Subsequently, a registration displacement table is generated according to the point-to-point coordinate difference between the first reference point column and the second reference point column. The coordinates of the car door object area and the suspected damage area in the first data are mapped to the second image to obtain the object confirmation area and the verification area. If the counter-evidence shooting action is a change in the incident light direction, the coordinates of the bright spot area are not used as registration constraints, but the bright spot area still participates in the mutual information sequence and coordinate difference sequence calculation as a non-entity element in S4.

[0020] S4. According to the suspicious area element type and target element label, extract entity elements and non-entity elements in the suspicious area and verification area to form entity coordinate sequence and non-entity coordinate sequence, and obtain coordinate comparison results. This implementation method is used to generate coordinate comparison results for S5 to read between the suspected area and the verification area. Its processing principle is as follows: First, extract the first target element point sequence and the second target element point sequence according to the target element markers. Then, use the reference area coordinates in the first data as the first reference area coordinates to generate the first coordinate sequence, and use the second reference area coordinates to generate the second coordinate sequence. Subsequently, divide the corresponding segments according to the point sequence order, calculate the first entropy sequence, the second entropy sequence, the joint entropy sequence, and the mutual information sequence. After mapping the first target element point sequence to the verification area through a registration displacement table, generate a reconstructed coordinate sequence based on the second reference area coordinates. Finally, subtract the reconstructed coordinate sequence from the second coordinate sequence point by point to obtain the coordinate difference sequence. This implementation process includes the following steps: S41 is used to determine the target element point sequence for comparison based on the target element marker, ensuring that the calculation object of S4 is consistent with the specific target element in S1 that participates in generating the conclusion to be effective; the edge end reads the coordinates of the suspicious area, the coordinates of the verification area, the coordinates of the first reference area, the coordinates of the second reference area, the element type of the suspicious area, the type of the target element, the target element marker, and the registration displacement table; the first target element point sequence is extracted in the suspicious area according to the target element marker, and the second target element point sequence is located in the verification area according to the target element type pointed to by the target element marker, the coordinates of the verification area, the point sequence order, and the coordinates of the second reference area; when the target element type is a boundary line, the first target element point sequence and the second target element point sequence are boundary points arranged in the order of line segment connection; when the target element type is a texture line, the first target element point sequence and the second target element point sequence are boundary points arranged in the order of line segment connection. The pixel column consists of texture points arranged in the order of texture extension; when the target feature type is a bright spot area, the first target feature point column and the second target feature point column consist of contour points arranged in the order of contour connection; when the target feature type is an occluded area and the occlusion relationship changes before entering S4, the edge end does not use the occluded area itself as the basis for confirming the entity feature, but extracts the boundary line, texture line or bright spot area exposed after the occlusion is removed in the verification area as the second target feature point column; if the second target feature point column is not extracted in the verification area, all the serial numbers of the first target feature point column are written into the missing serial number to be written; if only some corresponding points are missing in the verification area, the serial numbers of the first target feature point column that cannot find corresponding points in the second target feature point column are written into the missing serial number to be written, for S45 to write into the missing serial number field in the coordinate difference sequence; S42 is used to convert the first target element point list and the second target element point list to their respective reference coordinate bases, so that the subsequent entropy calculation and coordinate difference calculation have the same sequence number basis. The edge end uses the first reference area coordinates as the first coordinate base, and calculates the coordinates of each point in the first target element point list relative to the first reference area coordinates according to the reference origin and reference direction determined by S13, and generates the first coordinate sequence in the order of the point list. The edge end uses the second reference area coordinates as the second coordinate base. The reference origin of the second reference area is determined by the point in the second reference point list corresponding to the first reference origin, and the reference direction of the second reference area is determined by the point list direction in the second reference point list corresponding to the first reference direction. The edge end calculates the coordinates of each point in the second target element point list relative to the second reference area coordinates, and generates the second coordinate sequence in the order of the point list. If a certain number point in the first target element point list cannot be converted to the first coordinate sequence, or a certain number point in the second target element point list cannot be converted to the second coordinate sequence, the edge end writes the number to be written as a missing number, so that S45 can write it into the missing number field in the coordinate difference sequence. S43 is used to calculate the information distribution of the first and second coordinate sequences within corresponding segments, providing the first and second entropy sequences for subsequent mutual information sequence generation; the edge end divides the coordinate points with the same and continuous sequence number in the first and second coordinate sequences into corresponding segments according to the target element marking and point sequence order; when there is a missing sequence number to be written, the missing sequence number to be written is used as the segmentation breakpoint, and the continuous coordinate points before and after the missing sequence number to be written are divided into different corresponding segments; the edge end reads the gray value, edge direction, and texture direction of the pixels where the coordinate points are located in the corresponding segments of the first and second coordinate sequences respectively; the edge direction of the boundary point is obtained by the connection direction between the boundary point and the adjacent boundary point, and the edge direction of the contour points in the bright spot area is... The direction of the contour point is obtained by connecting the contour point to the adjacent contour point, and the direction of the texture point is obtained by extending the texture point to the adjacent texture point. When the target feature type is a bright spot area and the texture direction is an empty field, the empty field is not used as a texture direction value in the probability statistics. The number of occurrences of gray value, edge direction and texture direction in the corresponding segment is counted at the edge end. The number of occurrences is divided by the total number of readable pixels in the corresponding segment to obtain the probability value. The product of each probability value and the logarithm of each probability value is multiplied by the negative and then summed to obtain the first entropy sequence and the second entropy sequence. Gray value, edge direction or texture direction that does not appear in the corresponding segment is not written into the probability value. When there are no readable pixels in the corresponding segment, the corresponding segment number is written into the missing number to be written and the entropy value of the corresponding segment is not calculated. S44 is used to calculate the information correspondence between the first coordinate sequence and the second coordinate sequence, and generate a mutual information sequence; the edge end combines the gray value, edge direction, and texture direction of the pixels with the same index in the corresponding segment of the first coordinate sequence and the second coordinate sequence into a joint value record; this joint value record includes the gray value, edge direction, and texture direction of the same index coordinate point in the first coordinate sequence, and the gray value, edge direction, and texture direction of the same index coordinate point in the second coordinate sequence; empty fields in S43 that do not have probability values ​​written do not participate in the joint value record; when the same index coordinate point When a value field for statistical analysis is missing in either the first or second coordinate sequence, preventing the formation of a joint value record, the edge end writes the corresponding segment number into the missing segment number to be written, without calculating the joint entropy and mutual information of that corresponding segment. The edge end counts the occurrences of the joint value record, divides the occurrences of the joint value record by the total number of joint value records in the corresponding segment to obtain the joint probability value, and sums the products of each joint probability value and its logarithm, taking the opposite of each product, to obtain the joint entropy sequence. Subsequently, the first entropy sequence and the second entropy sequence are added segment by segment, and the joint entropy sequence is subtracted to obtain the mutual information sequence. S45 is used to write the information comparison results and spatial reconstruction results into the entity comparison field or non-entity comparison field, forming a coordinate comparison result for S5 to read; the edge end reads the registration displacement table, maps the first map coordinates of the first target feature point column to the verification area according to the registration displacement table, and converts them into relative coordinates based on the second reference area coordinates after mapping, to obtain the reconstructed coordinate sequence; the second coordinate sequence is also generated based on the second reference area coordinates, and the edge end subtracts the coordinates with the same sequence number in the reconstructed coordinate sequence from the coordinates with the coordinates with the same sequence number in the second coordinate sequence point by point to obtain the coordinate difference sequence; the coordinate difference sequence includes a coordinate difference field and a missing sequence number field: when both the reconstructed coordinates with the same sequence number and the second coordinates exist, the result of the point-by-point subtraction is written into the coordinate difference field; ... the reconstructed coordinates with the same sequence number and the second coordinates exist, the result of the point-by-point subtraction is written into the coordinate difference field. When a coordinate point or a second coordinate point is missing, the sequence number is written to the missing sequence number field; missing sequence numbers generated by S41 to S44 are also written to the missing sequence number field; the edge end performs field writing according to the suspected feature type: when the suspected feature type is an entity feature, the mutual information sequence and coordinate difference sequence are written to the entity comparison field; when the suspected feature type is a non-entity feature, the mutual information sequence and coordinate difference sequence are written to the non-entity comparison field; when multiple target feature markers exist, the edge end writes to the entity comparison field or non-entity comparison field for each target feature marker, and retains the correspondence between the target feature marker and the mutual information sequence and coordinate difference sequence; the entity comparison field and the non-entity comparison field are merged to generate the coordinate comparison result for S5 to read; Through this implementation method, S4 transforms the suspicious area in the first image and the verification area in the second image into field-based comparison results: the target element marker determines the first target element point sequence and the second target element point sequence; the first reference area coordinates and the second reference area coordinates determine the first coordinate sequence and the second coordinate sequence; the entropy sequence, joint entropy sequence, and mutual information sequence record the information correspondence of the target elements; the reconstructed coordinate sequence and coordinate difference sequence record the spatial correspondence of the target elements; the missing sequence number field records the sequence number that cannot form a corresponding point or corresponding segment; the writing of the entity comparison field and the non-entity comparison field is only completed based on the suspicious area element type in S1, and is not equivalent to final confirmation; S5 then controls and determines the audit conclusion, supplementary collection instruction, or partial recalculation instruction based on the mutual information sequence, coordinate difference sequence, and missing sequence number field in the entity comparison field and the non-entity comparison field; in practical applications: taking vehicle damage audit as an example, if the target element marker For the continuous scratch boundary line, the edge end extracts the first boundary point sequence in the suspected area of ​​the first image according to the line segment connection order, and extracts the second boundary point sequence in the verification area of ​​the second image according to the boundary line type, verification area coordinates, line segment connection order, and second reference area coordinates. The edge end generates a first coordinate sequence and a second coordinate sequence with the first reference area coordinates and the second reference area coordinates, respectively. The first entropy sequence, the second entropy sequence, the joint entropy sequence, and the mutual information sequence corresponding to the gray value, edge direction, and texture direction are calculated. Then, the first boundary point sequence is reconstructed to the verification area according to the registration displacement table and converted with the second reference area coordinates as the reference to generate a reconstructed coordinate sequence. Finally, the reconstructed coordinate sequence is subtracted from the second coordinate sequence point by point to generate a coordinate difference sequence. Since the suspected area element type of the continuous scratch boundary line is an entity element, the mutual information sequence and the coordinate difference sequence are written into the entity comparison field for S5 to determine whether the pending conclusion can be output as a confirmed audit conclusion.

[0021] S5. When the entity element forms a mutual information sequence and a coordinate difference sequence in the coordinate comparison result, the missing sequence number is empty, and there is no non-entity mutual information sequence and non-entity coordinate difference sequence, the pending conclusion is the confirmed review conclusion; when the mutual information sequence, coordinate difference sequence or missing sequence number of the entity element does not meet the output conditions, or when a non-entity mutual information sequence or non-entity coordinate difference sequence exists but there is no entity mutual information sequence and entity coordinate difference sequence, the confirmed review conclusion is stopped, and a supplementary collection instruction or a partial recalculation instruction is output. This implementation method controls whether a pending conclusion is approved as a confirmed review conclusion based on coordinate comparison results. The edge end first reads the entity comparison field and non-entity comparison field under the same target feature label, then determines whether the pending conclusion meets the output conditions based on the entity mutual information sequence, entity coordinate difference sequence, entity missing sequence number, and non-entity field status. When the entity field is complete and the non-entity field has not formed stable corresponding data under the same target feature label, the pending conclusion is written into the confirmed review conclusion. When the target feature comparison field is missing, the entity field is missing, the non-entity field is abnormally occupied, or the entity calculation field is incomplete, the output of the confirmed review conclusion is stopped, and a supplementary collection instruction or a partial recalculation instruction is generated based on the field source. This implementation process includes the following steps: S51 is used to extract the judgment field corresponding to the same target feature label from the coordinate comparison results, providing the field basis required for the conclusion admission of S52 and S53; the edge end reads the conclusion to be effective, target feature label, suspicious area coordinates, coordinate comparison results, and the second map acquisition requirements output by S2, and reads the entity comparison field and non-entity comparison field from the coordinate comparison results; then, using the target feature label as an index, it extracts the entity mutual information sequence, entity coordinate difference sequence, and entity missing number from the entity comparison field, and extracts the non-entity mutual information sequence, non-entity coordinate difference sequence, and non-entity missing number from the non-entity comparison field; entity missing number. Both the missing sequence number and the missing sequence number of the non-entity are derived from the missing sequence number field of the coordinate difference sequence in S4. The entity mutual information sequence, entity coordinate difference sequence, non-entity mutual information sequence and non-entity coordinate difference sequence all retain the correspondence with the target feature label. If neither the entity comparison field nor the non-entity comparison field exists under the same target feature label, the edge end records the missing status of the target feature comparison field and provides this status for S53 to read. If the target feature point sequence has been extracted in S4, but the entity mutual information sequence, entity coordinate difference sequence, non-entity mutual information sequence or non-entity coordinate difference sequence has not been generated, the edge end records the incomplete field status and provides this status for S53 to read. S52 is used to admit the pending conclusion as a confirmed review conclusion when the entity field is complete and there is no non-entity stable corresponding data under the same target feature label. The edge end reads the entity mutual information sequence, entity coordinate difference sequence, entity missing sequence number, non-entity mutual information sequence, and non-entity coordinate difference sequence extracted by S51. When the entity mutual information sequence exists, the entity coordinate difference sequence exists, the entity missing sequence number is empty, and neither the non-entity mutual information sequence nor the non-entity coordinate difference sequence exists, the edge end writes the pending conclusion into the confirmed review conclusion field and outputs it. Non-entity stable corresponding data refers to the existence of both the non-entity mutual information sequence and the non-entity coordinate difference sequence corresponding to the same target feature label. Furthermore, if there is no entity mutual information sequence or entity coordinate difference sequence corresponding to the same target element in the entity comparison field; if only one of the non-entity mutual information sequence or non-entity coordinate difference sequence exists, it will not be written into the "Determine Review Conclusion" field, and the non-entity abnormal interception status will be recorded by S53; if the non-entity element does not form a complete corresponding segment after the counter-evidence shooting action, or if the corresponding sequence number is written as the non-entity missing sequence number, it will be recorded as a non-entity weakened or disappeared state, and will not be used as non-entity stable corresponding data; if the entity mutual information sequence, entity coordinate difference sequence and entity missing sequence number do not meet the field combination of this step, the "Determine Review Conclusion" field will not be written at the edge end, and the process will be transferred to S53 to execute the stop output and processing instruction generation. S53 is used to prevent the pending conclusion from taking effect when the target feature comparison field is missing, the entity field is missing, non-entity fields are abnormally occupied, or the entity calculation field is incomplete, and generates a processing instruction after stopping output; the edge terminal reads the entity mutual information sequence, entity coordinate difference sequence, entity missing number, non-entity mutual information sequence, non-entity coordinate difference sequence, non-entity missing number, target feature comparison field missing status, and field incomplete status extracted by S51; when the target feature comparison field missing status exists, or any of the following is true: entity mutual information sequence does not exist, entity coordinate difference sequence does not exist, or entity missing number is not empty, the edge terminal stops outputting and confirms the review conclusion; when at least one of the non-entity mutual information sequence or non-entity coordinate difference sequence exists, and neither entity mutual information sequence nor entity coordinate difference sequence exists, the edge terminal records the non-entity abnormal interception status and stops outputting and confirms the review conclusion; when both the non-entity mutual information sequence and non-entity coordinate difference sequence exist, and the entity mutual information sequence and entity coordinate difference sequence do not exist ... entity mutual information sequence and entity coordinate difference sequence exist, and the entity coordinate difference sequence and entity coordinate difference sequence do not exist, the edge terminal records the non-entity abnormal interception status and stops outputting and confirms the review conclusion; when both the entity mutual information sequence and entity coordinate difference sequence exist, and the entity coordinate difference sequence and entity coordinate difference sequence do not exist, the edge terminal records the non-entity abnormal interception status and stops outputting and confirms When the volume coordinate difference sequence is not present, the edge end records the non-entity conflict state. The non-entity conflict state belongs to the complete non-entity corresponding state in the non-entity anomaly interception state. When the entity missing sequence number is not empty, the edge end reads the alignment field and action field from the second image acquisition requirements and writes the alignment field, action field, suspicious area coordinates and entity missing sequence number into the supplementary acquisition instruction. When the incomplete field state indicates that the entity mutual information sequence is not present and the entity coordinate difference sequence exists, or the entity mutual information sequence exists and the entity coordinate difference sequence is not present, the edge end writes the target element mark and suspicious area coordinates into the local recalculation instruction. The non-entity missing sequence number is used to record that the non-entity element has not formed a complete corresponding segment and is used to indicate the non-entity weakening or disappearance state. The non-entity missing sequence number does not trigger entity supplementary acquisition and is not used as a non-entity anomaly interception state on its own. The non-entity anomaly interception state is triggered when at least one of the non-entity mutual information sequence or the non-entity coordinate difference sequence exists and neither the entity mutual information sequence nor the entity coordinate difference sequence exists. Through this implementation method, S5 converts the coordinate comparison results generated by S4 into a conclusion admission output: when both the entity mutual information sequence and the entity coordinate difference sequence exist and the entity missing sequence number is empty, the entity comparison field under the same target element label meets the determination output condition; when neither the non-entity mutual information sequence nor the non-entity coordinate difference sequence exists, there is no non-entity stable corresponding data under the same target element label; when the target element comparison field is missing, the field is incomplete, the entity is missing, or the non-entity is abnormally occupied, the edge end first stops outputting the determination audit conclusion, and then generates a supplementary collection instruction or a local recalculation instruction according to the field source; in practical applications: taking vehicle damage audit as an example, if If both the entity mutual information sequence and the entity coordinate difference sequence corresponding to the continuous scratch boundary line exist, the entity missing sequence number is empty, and there are no non-entity mutual information sequence and non-entity coordinate difference sequence under the same target feature label, the edge end will write the pending conclusion "the door scratch is established" into the confirmed review conclusion and output it; if the entity missing sequence number of the continuous scratch boundary line is not empty, the edge end will read the alignment field and action field from the second image acquisition requirements, and write the suspicious area coordinates and entity missing sequence number into the supplementary acquisition instruction; if only the entity mutual information sequence is generated but the entity coordinate difference sequence is not generated, the edge end will write the target feature label and suspicious area coordinates of the continuous scratch boundary line into the local recalculation instruction.

[0022] Working principle: This solution first acquires the first image from the edge and generates a conclusion to be effective. Simultaneously, it identifies the object area, reference area, and suspicious area from the first image, recording the target element type, suspicious area element type, and target element marker. Then, based on the conclusion type and target element situation, it generates the second image acquisition requirements, obtaining the second image by changing the acquisition perspective, incident light direction, acquisition distance, or occlusion relationship. Next, it establishes a registration relationship using the reference area in the first image and the second reference area in the second image, mapping the suspicious area to the second image to form a verification area. Then, it extracts the corresponding target elements from the suspicious area and the verification area, calculates the mutual information sequence, coordinate difference sequence, and missing sequence number, forming entity comparison fields and non-entity comparison fields. Finally, it determines whether the conclusion to be effective can be output as a confirmed review conclusion based on these fields. If entity elements do not form a valid correspondence or non-entity elements still form interfering correspondences, the output stops and supplementary acquisition or partial recalculation is triggered. Taking vehicle damage review as an example, the system first takes a picture of the car door at the edge, identifies suspected scratches or bright spots, and selects stable structures such as the door waistline and corner points as reference areas. If the suspected target is the scratch boundary line, the system will require changing the acquisition angle or distance to take a second picture. If the suspected target is a reflective bright spot, the system will require changing the incident light direction to take a second picture. The system then aligns the second picture with the first picture according to the reference area and checks whether the suspected position can still form the same boundary line, texture line, or bright spot correspondence in the second picture. If the scratch boundary line remains corresponding in the two pictures and there is no non-physical interference such as bright spots, the damage is output as established. If the correspondence is missing, the field is incomplete, or the bright spot still occupies the suspected area, a definite conclusion is not given directly, but a re-photograph or partial recalculation is required.

[0023] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for controlling the consistency of image analysis results for business review, characterized in that, include: S1. Obtain the first image collected by the edge terminal for the review items. The first image is the conclusion generation image. The first image is segmented to determine the object area, reference area, and suspected area. The output includes a pending conclusion carrying the conclusion type and initial data. The initial data includes the object area coordinates, reference area coordinates, suspected area coordinates, target element type, suspected area element type, and target element label. The pending conclusion is a judgment result generated based on the first image but not output as the final review conclusion. The suspected area is a local region supporting the pending conclusion. The reference area is a stable structural region located within the object area, establishing relative coordinates between the suspected area and the suspected area. S2. Read the counter-evidence shooting action according to the conclusion type, target element type and doubtful area element type, write the doubtful area coordinates, reference area coordinates and counter-evidence shooting action into the acquisition command, and output the second image acquisition requirements; the counter-evidence shooting action changes the acquisition angle, incident light direction, acquisition distance or occlusion relationship, so that non-physical elements change the imaging position or display state, and the physical elements maintain the reference area coordinates. S3. Obtain the second image, which is a counter-evidence verification image collected according to the requirements of the second image collection; locate the second reference area, establish a registration relationship with the first reference point column and the second reference point column, and map the coordinates of the suspected area to the second image to obtain the verification area; S4. According to the suspicious area element type and target element label, extract entity elements and non-entity elements in the suspicious area and verification area to form entity coordinate sequence and non-entity coordinate sequence, and obtain coordinate comparison results. S5. When the entity element forms a mutual information sequence and a coordinate difference sequence in the coordinate comparison result, the missing sequence number is empty, and there is no non-entity mutual information sequence and non-entity coordinate difference sequence, the pending conclusion is the confirmed review conclusion. When the mutual information sequence, coordinate difference sequence or missing sequence number of the entity element does not meet the output conditions, or when there is a non-entity mutual information sequence or non-entity coordinate difference sequence but no entity mutual information sequence and entity coordinate difference sequence, the confirmed review conclusion is stopped, and a supplementary collection instruction or a partial recalculation instruction is output.

2. The image analysis result consistency control method for business review according to claim 1, characterized in that: S11. Obtain the first image collected by the edge end for the audit item, read the object category corresponding to the audit item, divide the outer contour of the audited object in the first image according to the object category, record the area enclosed by the outer contour as the object area, and extract the structural area with closed boundary, intersection point or continuous texture in the object area as the reference area. S12. Extract the local area that participates in generating the conclusion to be effective within the object area, and record the local area as the suspicious area. Extract the boundary line, texture line, bright spot area or occlusion area within the suspicious area. Record the boundary line, texture line, bright spot area or occlusion area that participates in generating the conclusion to be effective as the target feature. Output the coordinates of the suspicious area, the type of the target feature and the label of the target feature. S13. Using the closed boundary, intersection points, or continuous texture of the reference area as the coordinate reference, write the coordinates of the object area, the reference area, and the suspicious area into the same coordinate system; when the target feature type is a boundary line or texture line, and there is a connection relationship, relative position relationship, arrangement order, or connection direction between the target feature and the outer contour of the object area or the reference area point column, mark the target feature as a solid feature; when the target feature type is a bright spot area or an occluded area, mark the target feature as a non-solid feature, and output the conclusion to be effective and the first data.

3. The image analysis result consistency control method for business review according to claim 2, characterized in that: S2 includes: S21. Read the conclusion type of the conclusion to be effective and the coordinates of the suspicious area, the coordinates of the reference area, the target element type, the suspicious area element type and the target element marker in the first data. Write the conclusion type into the conclusion field, write the suspicious area element type into the element field, write the target element marker into the target field, write the target element type into the type field, and generate the counter-evidence items according to the conclusion field, element field, target field and type field. S22. Determine the counter-evidence shooting action based on the counter-evidence item: When the type field is bright spot area, write the change of incident light direction into the counter-evidence shooting action; when the type field is boundary line or texture line, write the change of acquisition view and acquisition distance into the counter-evidence shooting action; when the type field is occlusion area, switch to the occlusion relationship change action generation process. S23. When the type field is bright spot area, boundary line or texture line, the acquisition boundary is formed by the coordinates of the suspicious area and the coordinates of the reference area. The acquisition boundary is written into the framing field, the counter-evidence shooting action is written into the action field, the coordinates of the reference area in the first data are written into the alignment field, the coordinates of the suspicious area are written into the verification field, the acquisition command is generated, and the acquisition command is output as the acquisition requirements of the second image.

4. The image analysis result consistency control method for business review according to claim 3, characterized in that: S2 further includes: S24. Read the coordinates of the suspected area, the coordinates of the reference area in the first data, and the target element marker. Read the occlusion area segmentation result corresponding to the target element marker within the coordinates of the suspected area, and obtain the contour points of the occlusion area through grayscale edge detection, segmentation mask boundary extraction, or connected region outer contour extraction. Generate the contour center coordinates based on the horizontal and vertical coordinates of the contour points of the occlusion area. Record the direction opposite to the direction of the reference origin contained in the reference area coordinates in the first data as the occlusion direction. Write the occlusion direction into the occlusion field. Generate an occlusion relationship change action based on the occlusion field. Write the occlusion relationship change action into the action field. Write the coordinates of the suspected area into the verification field. Write the coordinates of the reference area in the first data into the alignment field. Generate an occlusion counter-evidence acquisition instruction. Output the occlusion counter-evidence acquisition instruction as the second image acquisition requirement.

5. The image analysis result consistency control method for business review according to claim 4, characterized in that: S3 includes: S31. Obtain the second image collected according to the second image collection requirements; read the object area coordinates, reference area coordinates, and suspicious area coordinates from the first data; and read the counter-evidence shooting action from the second image collection requirements. In the second image, extract closed boundaries, intersection points, or continuous textures according to the reference type corresponding to the reference area of ​​the first image, and generate a second candidate point column according to the connection, adjacency, or extension relationship. In the reference area of ​​the first image, generate a first reference point column according to the same reference type, and mark the coordinates of the area where the first reference point column is located as the coordinates of the first reference area. The point column order is the sequence number relationship formed by the connection, adjacency, or extension relationship of each point in the point column. Record the point column in the second candidate point column that has the same point column order as the first reference point column as the second reference point column, mark the area where the second reference point column is located as the second reference area, and mark the coordinates of the area where the second reference point column is located as the coordinates of the second reference area. S32. Establish point pairs according to the points in the first reference point column and the second reference point column with the same order. Write the coordinate difference of each point pair into the registration displacement table. Based on the registration displacement table, map the object area coordinates and doubt area coordinates in the first data point by point to the second figure to generate the object area mapping area and the doubt area mapping area.

6. The image analysis result consistency control method for business review according to claim 5, characterized in that: S3 further includes: S33. The overlapping area between the object area mapping region and the same audited object area in the second figure is recorded as the object confirmation area. The overlapping area between the suspicious area mapping region and the object confirmation region is recorded as the initial verification area. The retained objects in the initial verification area are adjusted according to the counter-evidence shooting action: when the incident light direction changes, the corresponding reference area point sequence is retained but the corresponding bright spot area coordinates are not retained; when the acquisition angle or acquisition distance changes, the corresponding reference area point sequence is retained but the corresponding suspicious area scale is not retained; when the occlusion relationship changes, the corresponding reference area point sequence is retained but the corresponding occlusion contour is not retained. The adjusted initial verification area is recorded as the verification area, and the verification area coordinates and registration displacement table are output.

7. The image analysis result consistency control method for business review according to claim 6, characterized in that: S4 includes: S41. Read the coordinates of the suspicious area, the coordinates of the verification area, the coordinates of the reference area and the second reference area in the first data, the suspicious area feature type, the target feature type, the target feature marker, and the registration displacement table. Extract the first target feature point column in the suspicious area according to the target feature marker, and extract the second target feature point column in the verification area. When the target feature type is a boundary line, the first target feature point column and the second target feature point column are boundary points arranged in the order of line segment connection. When the target feature type is a texture line, the first target feature point column and the second target feature point column are texture points arranged in the order of texture extension. When the target feature type is a bright spot area, the first target feature point column and the second target feature point column are contour points arranged in the order of contour connection. S42. Using the coordinates of the reference area in the first data as a reference, calculate the coordinates of each point in the first target element point list relative to the coordinates of the reference area in the first data, and generate a first coordinate sequence; using the coordinates of the second reference area as a reference, calculate the coordinates of each point in the second target element point list relative to the coordinates of the second reference area, and generate a second coordinate sequence. S43. According to the target element marking and point sequence, divide the consecutive coordinate points with the same sequence number in the first coordinate sequence and the second coordinate sequence into corresponding segments. Read the gray value, edge direction and texture direction of the pixels where the coordinate points are located in the corresponding segments of the first coordinate sequence and the corresponding segments of the second coordinate sequence respectively. Count the occurrence of gray value, edge direction and texture direction. Divide the occurrence by the total number of pixels in the corresponding segment to obtain the probability value. Then sum the products of each probability value and the logarithm of each probability value after taking the opposite number to obtain the first entropy sequence and the second entropy sequence.

8. The image analysis result consistency control method for business review according to claim 7, characterized in that: S4 further includes: S44. Take the gray value, edge direction, and texture direction of the pixels with the same index in the corresponding segment of the first coordinate sequence and the second coordinate sequence to form a joint value record. Count the number of occurrences of the joint value record. Divide the number of occurrences of the joint value record by the total number of joint value records in the corresponding segment to obtain the joint probability value. Take the negative of the product of each joint probability value and the logarithm of each joint probability value and sum them to obtain the joint entropy sequence. Add the first entropy sequence and the second entropy sequence segment by segment and then subtract the joint entropy sequence to obtain the mutual information sequence. S45. Map the first map coordinates of the first target feature point list to the verification area according to the registration displacement table, and convert them to the second reference area coordinates to obtain the reconstructed coordinate sequence. Subtract the coordinate points with the same sequence number in the reconstructed coordinate sequence from the coordinate points with the same sequence number in the second coordinate sequence to obtain the coordinate difference sequence. If the coordinate points with the same sequence number are missing, write the missing sequence number into the coordinate difference sequence. If the suspected feature type is an entity feature, write the mutual information sequence and the coordinate difference sequence into the entity comparison field. If the suspected feature type is a non-entity feature, write the mutual information sequence and the coordinate difference sequence into the non-entity comparison field. Merge the entity comparison field and the non-entity comparison field to generate the coordinate comparison result.

9. The image analysis result consistency control method for business review according to claim 8, characterized in that: S5 includes: S51. Read the pending conclusion, the coordinates of the suspicious area, the target element marker and the coordinate comparison result. Read the entity comparison field and the non-entity comparison field from the coordinate comparison result. Extract the entity mutual information sequence, entity coordinate difference sequence and entity missing number corresponding to the target element marker from the entity comparison field. Extract the non-entity mutual information sequence, non-entity coordinate difference sequence and non-entity missing number corresponding to the target element marker from the non-entity comparison field. S52. When both the entity mutual information sequence and the entity coordinate difference sequence exist, the entity missing sequence number is empty, and neither the non-entity mutual information sequence nor the non-entity coordinate difference sequence exists, the pending conclusion is written into the confirmed audit conclusion and output.

10. The image analysis result consistency control method for business review according to claim 9, characterized in that: The S5 also includes: S53. When any of the following conditions are met: the entity mutual information sequence does not exist, the entity coordinate difference sequence does not exist, or the entity missing sequence number is not empty, stop outputting the confirmed audit conclusion; when a non-entity mutual information sequence or a non-entity coordinate difference sequence exists, and neither the entity mutual information sequence nor the entity coordinate difference sequence exists, stop outputting the confirmed audit conclusion; when the entity missing sequence number is not empty, write the suspicious area coordinates and the entity missing sequence number into the supplementary collection instruction; when the entity mutual information sequence does not exist and the entity coordinate difference sequence exists, or when the entity mutual information sequence exists and the entity coordinate difference sequence does not exist, write the target element mark and the suspicious area coordinates into the local recalculation instruction.

Images