Orthodontic profile image feature extraction method based on computer vision

Abstract

The application relates to the field of computer vision and medical image processing technology, and discloses an orthodontic side appearance image feature extraction method based on computer vision, which comprises the following steps: acquiring an orthodontic side appearance image, and performing standardization pretreatment on the orthodontic side appearance image to obtain a standardized image; scanning the standardized image along a preset reading direction; determining candidate morphological feature segments according to side appearance outer edge position changes, pixel gray scale changes and connected region ranges; performing rule screening on the candidate morphological feature segments, and deleting candidate morphological feature segments that do not satisfy the rules; checking the candidate morphological feature segments reserved after screening according to the consistency of starting positions, ending positions and lengths, and determining side appearance feature segments; generating sequential position codes according to the arrangement order and interval relationship of the side appearance feature segments in the reading direction, and outputting side appearance feature data. The application scheme can improve the stability and consistency of side appearance feature recognition, and has good application value.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and medical image processing technology, specifically to a method for extracting features from orthodontic lateral profile images based on computer vision. Background Technology

[0002] As orthodontics evolves from simply correcting occlusion to focusing on facial aesthetics, soft tissue coordination, and personalized treatment design, orthodontic facial profile information plays an increasingly prominent role in clinical diagnosis, treatment planning, and efficacy evaluation. Currently, the assessment of a patient's facial shape and profile aesthetics largely relies on physician observation, patient subjective feelings, or coarse classification based on a limited number of landmarks. For example, Chinese patent application CN118447548A points out that existing orthodontic facial aesthetic classification is primarily based on qualitative judgments from physicians and patients. Traditional two-dimensional facial measurements are insufficient to meet the accuracy requirements of facial classification measurements, and even with three-dimensional facial classification, the accuracy and standards still fall short of the demands for standardization, objectivity, and quantification. It is evident that existing methods for representing facial profile contours often remain at a coarse-grained level, such as "straight, convex, or concave," making it difficult to stably describe the continuous morphological changes of the outer edge of the profile in local segments, the start and end positions of the boundaries, and the sequential relationships between adjacent segments. This results in insufficient consistency in judgment results among different operators and under different acquisition conditions, poor repeatability, and difficulty in directly forming structured data that is easy for computers to process further.

[0003] On the other hand, existing technologies also include schemes for automatic orthodontic detection and analysis using CBCT (cone-beam computed tomography), 3D point clouds, and deep learning networks. For example, Chinese patent application CN117830317A discloses a technical approach that extracts tooth axis rotation transformation information through CBCT image segmentation, image transformation, point cloud processing, and point-by-point prediction networks. This document also explicitly points out that CBCT images generally suffer from low contrast, indistinct jawbone edges compared to the background, noise due to radiation dose limitations, and difficulty in distinguishing details, thus affecting the accuracy of segmentation algorithms. Furthermore, existing tooth axis markings often require manual work by doctors, lack clear mathematical definitions for orthodontic features, and relying solely on global geometric direct regression is insufficient to detect all features. Therefore, while existing technologies relying on CBCT, 3D reconstruction, and deep neural networks can provide relatively rich anatomical and spatial information, they often suffer from drawbacks in widespread clinical applications, including strong equipment dependence, high data acquisition costs, high dependence on annotation quality and training samples, and relatively complex processing procedures. For applications that only use ordinary orthodontic profile images as input, if they still mainly rely on heavy 3D detection or large-scale learning models, problems such as unstable feature localization, false positives and false negatives, and insufficient interpretability of results are likely to occur under conditions of pose changes, scale differences, grayscale fluctuations, and local weak boundaries.

[0004] As can be seen from the existing technologies described above, one type of approach focuses on facial shape classification but lacks a refined extraction mechanism for local morphological segments in lateral profile images that can be directly used for quantitative analysis. Another type of approach focuses on 3D detection and deep feature learning under CBCT or point cloud conditions but has not established a lightweight, interpretable, and easily engineering-implemented rule-based feature extraction process for ordinary lateral profile images. Especially in orthodontic lateral profile image processing, without unified scanning rules for pose correction, scale normalization, and grayscale normalization, without joint constraints on changes in the position of the lateral profile's outer edge, pixel grayscale changes, and the range of connected regions, and without step-by-step screening of candidate segment continuity, positional correspondence, and consistency between adjacent scans, the resulting feature segments often exhibit significant boundary drift, inconsistent segment definitions, and ambiguous relationships between adjacent segments. This makes it difficult to meet the requirements of subsequent lateral profile quantitative assessment, case-to-case comparison, treatment course tracking, and intelligent analysis interfaces for stability, repeatability, and structured data output. Therefore, we have invented a computer vision-based orthodontic lateral profile image feature extraction method that solves the above technical problems. Summary of the Invention

[0005] This invention provides a computer vision-based method for extracting orthodontic profile features from images, which can transform profile contour information in the original image into structured feature information with clear interval attributes, thereby improving the stability and consistency of profile feature recognition.

[0006] Therefore, the present invention provides the following technical solution: A method for extracting features from orthodontic lateral profile images based on computer vision, the method comprising: Step 1: Obtain orthodontic profile images and perform standardized preprocessing on the orthodontic profile images to obtain standardized images; Step 2: Scan the standardized image along the preset reading direction, and determine candidate morphological feature segments based on changes in the outer edge position of the profile, changes in pixel grayscale, and the range of connected regions. Step 3: Perform rule-based filtering on candidate morphological feature segments and delete candidate morphological feature segments that do not meet the rules; Step 4: Verify the consistency of the candidate morphological feature fragments retained after screening according to the start position, end position and length to determine the side profile feature fragments. Step 5: Generate sequential position codes based on the arrangement order and interval relationship of the profile feature fragments in the preset reading direction, and output the profile feature data.

[0007] Optionally, in step 2, the preset reading direction is preset according to the image coordinate system or the unfolding direction of the side profile. Based on the joint determination of three dimensions—the change in the position of the outer edge of the side profile, the change in pixel grayscale, and the range of the connected region—multiple candidate morphological feature segments are determined. The candidate morphological feature segments are local intervals with start and end positions in the preset reading direction. The local intervals correspond to morphological changes in the side profile that may have diagnostic or structural significance.

[0008] Optionally, in step 2, during scanning, a forward scan is performed along the preset reading direction, and a reverse scan is performed in the opposite direction to the preset reading direction. In both the forward and reverse scans, candidate morphological feature segments are obtained based on a joint determination of changes in the side profile outer edge position, pixel grayscale changes, and the range of connected regions. The start and end positions of the candidate morphological feature segments obtained by the forward scan are matched with the end and start positions of the candidate morphological feature segments obtained by the reverse scan, respectively. When the position difference between the corresponding positions is within a preset position difference threshold range, the match is successful, and the successfully matched segment is retained as a candidate morphological feature segment. The matching includes: when there are multiple candidate morphological feature segments obtained by the reverse scan corresponding to the same candidate morphological feature segment obtained by the forward scan, the sum of the absolute values ​​of the position differences between the corresponding positions of each candidate morphological feature segment obtained by the reverse scan is calculated, and the segments are sorted in ascending order of the sum of the absolute values ​​of the position differences. The candidate morphological feature segment obtained by the reverse scan with the smallest sum of the absolute values ​​of the position differences is determined as the segment to be matched with the candidate morphological feature segment obtained by the forward scan.

[0009] Optionally, in step 2, after the position difference between corresponding positions is within a preset position difference threshold range, a first position interval is determined based on the candidate morphological feature segments obtained by forward scanning, and a second position interval is determined based on the candidate morphological feature segments obtained by reverse scanning; when the ratio of the overlap interval length to the merged interval length of the first position interval and the second position interval reaches a preset overlap ratio threshold, it is determined that the candidate morphological feature segments obtained by forward scanning and the candidate morphological feature segments obtained by reverse scanning satisfy the interval consistency condition. When the interval consistency condition is met, a first signed position difference is calculated between the starting position of the candidate morphological feature segment obtained by forward scanning and the corresponding starting position of the candidate morphological feature segment obtained by reverse scanning after unified coordinate conversion, and a second signed position difference is calculated between the ending position of the candidate morphological feature segment obtained by forward scanning and the corresponding ending position of the candidate morphological feature segment obtained by reverse scanning after unified coordinate conversion. When the signs of the first signed position difference and the second signed position difference are consistent and their absolute values ​​are within a preset offset threshold range, the arithmetic mean of the starting position obtained by forward scanning and the corresponding starting position after unified coordinate conversion is determined as the starting position of the candidate morphological feature segment, and the arithmetic mean of the ending position obtained by forward scanning and the corresponding ending position after unified coordinate conversion is determined as the ending position of the candidate morphological feature segment. The corresponding starting position after unified coordinate conversion and the corresponding ending position after unified coordinate conversion are respectively obtained by converting the ending position and the starting position recorded by reverse scanning in the scanning order.

[0010] Optionally, in step 3, candidate morphological feature segments are filtered according to rules based on the continuity of the candidate morphological feature segments in the preset reading direction, the positional correspondence between pixel grayscale changes and outer edge position changes, and the consistency of adjacent scanning results. Specifically, this includes: Step 31: Determine the starting point of the outer edge position change, the ending point of the outer edge position change, the starting point of the pixel grayscale change, and the boundary of the connected region for each candidate morphological feature segment. Step 32: Determine the sequential relationship between pixel grayscale change and outer edge position change, as well as the closure relationship of the outer edge position change region. When the starting point of the outer edge position change is located before or coincides with the starting point of the pixel grayscale change in the preset reading direction, or when the ending point of the outer edge position change is not located within the connected region, delete the corresponding candidate morphological feature fragment. Step 33: Determine the positional correspondence between pixel grayscale changes and outer edge position changes. Calculate a first positional difference from the starting point of the pixel grayscale change to the starting point of the outer edge position change, and a second positional difference from the ending point of the outer edge position change to the boundary of the connected region, respectively, along the reading direction. When the first positional difference does not reach a preset leading threshold, or the second positional difference does not reach a preset closing threshold, delete the corresponding candidate morphological feature segment. Determine the outer edge change interval using the starting point and ending point of the outer edge position change as boundaries, and determine the constraint interval using the starting point of the pixel grayscale change and the boundary of the connected region, respectively. When the outer edge change interval is not completely contained by the constraint interval, delete the corresponding candidate morphological feature segment. Step 34: Filter based on the consistency of adjacent scan results. If a candidate morphological feature segment is detected only at a single scan position and cannot be supported in adjacent scan positions, or if the range of the start position, the range of the end position, or the relative range of the length of the same local structure in multiple adjacent scan results exceeds the corresponding preset fluctuation threshold, then the candidate morphological feature segment is deemed to have insufficient stability and is deleted.

[0011] Optionally, in step 4, when the starting position of the candidate morphological feature fragment retained after screening remains stable within the allowable fluctuation range, the ending position remains stable within the allowable fluctuation range, and the length of the candidate morphological feature fragment does not show abnormal expansion or abnormal contraction, the candidate morphological feature fragment is determined as a valid profile feature fragment. The length of the candidate morphological feature fragment is determined by the range between the ending position and the starting position of the candidate morphological feature fragment. The allowable fluctuation range is preset according to the image resolution, standardization scale, or sample statistical results.

[0012] Optionally, in step 5, the start position, end position, and length of each profile feature segment in the preset reading direction are determined; sequential position codes are generated according to the positional interval relationship and interval relationship between adjacent profile feature segments; wherein, the interval relationship includes separation, connection, overlap, and inclusion.

[0013] Optionally, in step 5, a hierarchical sequential position code is generated based on the position difference between two adjacent profile feature segments in the preset reading direction, the ratio of the length of the preceding profile feature segment to the length of the following profile feature segment, and the interval relationship; for different adjacent profile feature segment pairs whose interval relationship belongs to the same category, the hierarchical sequential position code is further divided based on the position difference and the length ratio.

[0014] A computer vision-based orthodontic lateral profile image feature extraction system, the system comprising: The image preprocessing unit acquires orthodontic profile images and performs standardized preprocessing on the orthodontic profile images to obtain standardized images. The candidate morphological feature segment determination unit scans the standardized image along a preset reading direction and determines candidate morphological feature segments based on changes in the outer edge position of the profile, changes in pixel grayscale, and the range of connected regions. The candidate morphological feature segment filtering unit performs rule-based filtering on candidate morphological feature segments and deletes candidate morphological feature segments that do not meet the rules. The feature fragment verification unit verifies the candidate morphological feature fragments retained after screening according to the consistency of the start position, end position and length, and determines the profile feature fragments. The position encoding unit generates sequential position codes based on the arrangement order and interval relationship of the profile feature fragments in the reading direction, and outputs the profile feature data.

[0015] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to perform the steps of the computer vision-based orthodontic profile image feature extraction method.

[0016] The present invention provides a computer vision-based method for extracting orthodontic lateral profile image features. By first performing pose correction, scale normalization, and grayscale normalization on the lateral profile image, and then scanning along a preset reading direction, and combining the changes in the position of the outer edge of the lateral profile, the changes in pixel grayscale, and the range of connected regions to determine candidate morphological feature segments, the method can transform the lateral profile contour information in the original image into structured feature information with clear interval attributes. This reduces the interference caused by the shooting angle, imaging ratio, and brightness fluctuations on the extraction results, and improves the stability and consistency of lateral profile feature recognition.

[0017] The present invention further filters candidate segments according to the continuity of candidate morphological feature segments in the reading direction, the positional correspondence between pixel grayscale changes and outer edge position changes, and the consistency of adjacent scanning results. It also verifies the consistency of start position, end position and length, which can effectively eliminate noise interference, false boundaries and unstable segments, reduce false detection and missed detection, and make the determined side profile feature segments more reliable and repeatable.

[0018] Furthermore, this invention generates sequential position codes based on the arrangement order and interval relationship of lateral feature segments in the reading direction, and outputs lateral feature data. This allows the extraction results to not only characterize the local morphology of the lateral face, but also reflect the relative positional relationship between each feature segment, which facilitates subsequent orthodontic diagnostic assistance, case comparison analysis, efficacy tracking and evaluation, and related intelligent processing, and has good application value. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a computer vision-based orthodontic profile image feature extraction method in a specific embodiment of the present invention; Figure 2 This is a schematic diagram of the preferred bidirectional scanning matching of candidate fragments in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the overlap ratio consistency determination in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the same-direction offset boundary correction in Embodiment 1 of the present invention; Figure 5 This is a line graph showing the quantitative screening of leading and closing relationships in Embodiment 2 of the present invention. Figure 6 This is a schematic diagram illustrating the relationship between the outer edge variation range and the constraint range in Embodiment 2 of the present invention; Figure 7 This is a schematic diagram of the structure of a computer vision-based orthodontic profile image feature extraction system in a specific embodiment of the present invention. Detailed Implementation

[0021] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0023] like Figure 1 As shown, Figure 1 This is a flowchart of a computer vision-based orthodontic lateral profile image feature extraction method according to a specific embodiment of the present invention. The computer vision-based orthodontic lateral profile image feature extraction method includes: Step 1: Obtain orthodontic profile images, and perform pose correction, scale normalization, and grayscale normalization on the orthodontic profile images to obtain standardized images.

[0024] First, orthodontic lateral profile images are acquired. These images can be obtained using lateral profile imaging devices, oral imaging devices, or other imaging devices suitable for acquiring lateral facial images. To improve the stability and comparability of subsequent feature extraction results, the orthodontic lateral profile images undergo standardization preprocessing. Specifically, an image coordinate system for the orthodontic lateral profile images is established, and the region containing the lateral profile subject is determined. This region can be obtained through foreground segmentation, contour extraction, threshold separation, or manual pre-annotation. Next, pose correction is performed on the orthodontic lateral profile images to ensure that the orientation of the lateral profile subject in the image meets a uniform angular reference. Pose correction can be performed based on pre-defined reference points, reference lines, the subject's central axis, or the direction of its circumscribed contour to reduce the impact of head tilt, rotation, or slight sway on feature extraction. Finally, the pose-corrected images are scaled to normalize the image, mapping lateral profile images formed under different acquisition distances, resolutions, or individual conditions to a uniform scale range. Scale normalization can be performed by proportional scaling based on the circumscribed rectangle size of the lateral profile subject, reference distance, or target size. Next, grayscale normalization is performed to keep the image brightness and contrast distribution within a preset range, thereby reducing the interference of differences in acquisition lighting, device response, and local shadows on subsequent analysis. Grayscale normalization can be achieved through linear grayscale stretching, grayscale range compression, piecewise mapping, or histogram adjustment. In one embodiment, denoising processing can also be performed before or after grayscale normalization to remove isolated noise points and local texture fluctuations. After the above processing, a standardized image is obtained. The preset range, target size, angle reference, and related thresholds can be predetermined through historical sample statistics, manual annotation results, or pre-experimental parameter tuning, and should be kept consistent in the same batch of image processing.

[0025] Step 2: After obtaining the standardized image, scan the standardized image along the preset reading direction, and determine the candidate morphological feature segments based on the changes in the position of the outer edge of the profile, the changes in pixel grayscale, and the range of connected regions.

[0026] The preset reading direction can be pre-set according to the image coordinate system or the unfolding direction of the profile contour, thereby ensuring that the scanning rules are consistent between different images. The preferred reading direction is a row-by-row or column-by-column direction in the image coordinate system. Each scanning position corresponds to a scan line or a scan band with a preset width, which can be one pixel or multiple pixels wide. During the scanning process, contour change information, grayscale change information, and region connectivity information are extracted for each scanning position. Specifically, the position of the outer edge of the profile is first determined at the current scanning position. The position of the outer edge can be defined as the boundary between the foreground and background regions at the current scanning position, or as the coordinate value of the outer edge contour point perpendicular to the reading direction. Subsequently, the positional change of the outer edge of the profile in the reading direction is detected. When the outer edge contour shows significant shifts, turns, or extensions between adjacent scanning positions, it can be considered that there are potential morphological features at that position. Significant shifts, turns, or extensions can be determined by comparing the contour coordinate difference between adjacent scanning positions, contour direction changes, local curvature changes, or contour extension length changes with preset change conditions. Simultaneously, the pixel grayscale of the corresponding region is analyzed. Analysis can be based on individual pixel grayscale values, local window grayscale mean, local window grayscale extreme values, grayscale difference results, or grayscale change trends. When pixel grayscale shows gradient changes, abrupt grayscale changes, or a transition trend from foreground to background along the reading direction, this change can be used as an auxiliary criterion. Further, combined with the connected region range, it is determined whether the aforementioned outer edge change region is within a reasonable foreground connected region and whether it maintains a spatial relationship with the main profile. The connected region range can be determined by the foreground connected domain containing the outer edge change region, with the start and end boundaries of the projection of this connected domain in the reading direction as the range boundary. Based on the joint determination of the three dimensions of profile outer edge position change, pixel grayscale change, and connected region range, multiple candidate morphological feature segments are identified. Candidate morphological feature segments can be understood as local intervals with start and end positions in the reading direction, corresponding to morphological changes in the profile contour that may have diagnostic or structural significance. Both the start and end positions are recorded in a unified image coordinate system for subsequent position comparison and interval calculation.

[0027] In one feasible joint determination method, the outer edge change amount E(k) and grayscale change amount G(k) are calculated for each scanning position along a preset reading direction. The outer edge change amount E(k) can be determined by the outer edge coordinate difference, contour direction difference, local curvature difference, or a weighted combination thereof between adjacent scanning positions. When E(k) reaches a preset outer edge change threshold, and the number of scanning positions that continuously reach this threshold reaches a preset number of consecutive values, the continuous interval is initially determined as the outer edge change interval, and the outer edge position change start point xos and outer edge position change end point xoe are recorded. The grayscale change amount G(k) can be determined by the grayscale gradient, the difference in local window mean, or the foreground-background transition amplitude. When a position where G(k) reaches the preset grayscale change threshold is detected within a preset associated window before the outer edge position change start point, the position closest to the outer edge position change start point and satisfying the condition is recorded as the pixel grayscale change start point xg.

[0028] Further, a foreground connected region containing the outer edge change interval is determined, and the projection range of this foreground connected region in a preset reading direction is denoted as [xcs, xce]. When the outer edge change interval [xos, xoe] is located within the projection range [xcs, xce], the pixel grayscale change start point xg is no later than the outer edge position change start point xos, and the distance between the two does not exceed a preset association window, and the outer edge change, grayscale change, and foreground connected region correspond to the same local region, the preliminary joint conditions of the three dimensions are determined to be met, and a candidate morphological feature fragment is generated. The candidate morphological feature fragment is at least associated with records {xos, xoe, xg, xcs, xce} and the corresponding scanning position for subsequent matching and filtering.

[0029] The aforementioned joint determination is used to generate initial candidate segments under relatively lenient conditions; both forward and reverse scans execute the joint determination. Subsequently, the corresponding boundary difference and interval overlap ratio of the bidirectional scans are used to perform cross-directional consistency verification on the initial candidate segments, and the boundaries are corrected when the conditions are met; then, the starting point of pixel grayscale change, the starting and ending points of outer edge position change, and the boundary of connected region range are used to filter for leading relationships, closing relationships, and interval inclusion relationships. Therefore, bidirectional scan matching and rule filtering do not redefine features without considering the three dimensions, but rather perform cross-validation and progressive tightening on the initial candidate segments generated from the three dimensions and carrying corresponding positional parameters.

[0030] Step 3: Based on the continuity of the candidate morphological feature segments in the reading direction, the positional correspondence between the pixel grayscale change and the outer edge position change, and the consistency of adjacent scanning results, the candidate morphological feature segments are filtered according to rules, and candidate morphological feature segments that do not meet the rules are deleted.

[0031] After obtaining candidate morphological feature fragments, rule-based filtering is performed to eliminate noise, spurious features, and abnormal fragments caused by local occlusion or imaging fluctuations. First, candidate morphological feature fragments representing the same local structure in adjacent scan positions are grouped together based on conditions such as overlapping positions, close center positions, or boundary distances less than a preset merging threshold. If a candidate morphological feature fragment is detected only in a single scan position, or if the number of scan positions supporting it within a preset neighborhood containing that scan position is less than a preset number of supported positions, it is considered to have insufficient continuity and is deleted. Then, the sequential relationship and positional correspondence between pixel grayscale changes and outer edge position changes are determined: the starting point of the pixel grayscale change should be before the starting point of the outer edge position change, and the ending point of the outer edge position change should be within the effective range of the corresponding foreground connected region, further satisfying the leading distance, closing distance, and interval inclusion conditions. Through the joint filtering of the above continuity rules, positional correspondence rules, and adjacent consistency rules, candidate morphological feature fragments that better match the actual side profile structure are retained.

[0032] For the i-th candidate morphological feature segment in the same group, let its start position, end position, and length be respectively... , and ,in The range of the initial position can be calculated separately. Termination position range and the relative range of length ,in This represents the median length of the candidate segments in the same group. , or If any of the following exceeds the corresponding preset fluctuation threshold, the change in the starting position, ending position, or length can be considered excessive. The preset fluctuation threshold can be determined according to the resolution of the standardized image, the width of the scan band, and the statistical distribution of historical samples; for example, in a non-limiting implementation where the height of the profile subject is normalized to 512 pixels, the starting position threshold and the ending position threshold can be set to 3 pixels each, the length relative range threshold can be set to 15%, and the same segment can be required to be supported by at least 2 scan positions out of 3 consecutive scan positions.

[0033] Step 4: For the candidate morphological feature fragments retained after screening, further verification is performed based on the consistency of the start position, end position, and length to determine the final profile feature fragments. Specifically, the retained fragments representing the same local structure in multiple adjacent scan results are first classified according to their positional overlap, boundary proximity, or center proximity. Then, the distribution of the start position, end position, and length of each category of retained fragments in multiple adjacent scan results is statistically analyzed and compared with preset consistency conditions. When the start position of a fragment remains stable within the allowable fluctuation range, the end position remains stable within the allowable fluctuation range, and the fragment length does not show abnormal expansion or contraction, the fragment is determined as a valid profile feature fragment. Conversely, if a fragment passes the initial screening, but its boundary position or length fluctuates significantly in consecutive scan results, it indicates that its boundary is not yet stable and is insufficient as the final feature output; it can be excluded or re-merged. The length can be determined by the range between the end position and the start position, and the allowable fluctuation range can be preset according to image resolution, standardization scale, or sample statistical results.

[0034] Step 3 focuses on deleting candidate segments that are obviously discontinuous or do not meet structural relationships, while step 4 performs final consistency verification on the segments in the same group that have passed the screening. Step 4 can use the same or more stringent threshold as the aforementioned range index, and can use the median or weighted average of the starting positions of the segments in the same group as the final starting position, and the median or weighted average of the ending positions of the segments in the same group as the final ending position, in order to reduce the impact of outliers at individual scan positions on the final boundary.

[0035] Step 5: After determining the profile feature segments, generate a sequential position code based on the arrangement order and interval relationships of the profile feature segments in the reading direction. Specifically, first, sort the profile feature segments according to their spatial order in the reading direction and assign a segment sequence identifier to each profile feature segment; then, determine the interval relationships (separation, connection, overlap, or inclusion) for each pair of adjacent profile feature segments and assign a relationship identifier to that pair of adjacent profile feature segments. The segment sequence identifier and the relationship identifier of adjacent segment pairs together constitute the sequential position code. The sequential position code can be stored as a field associated with the corresponding profile feature segment, or as a record associated with the corresponding pair of adjacent profile feature segments; regardless of the storage method used, the encoded object clearly points to a profile feature segment or a pair of adjacent profile feature segments. The sequential position code can adopt integer encoding, string encoding, binary encoding, array encoding, or field combination encoding, and its content includes at least the segment sequence number and interval relationship identifier; when necessary, it can further include the start position, end position, length, adjacent segment identifier, or position interval level. Finally, the profile feature fragments and their corresponding sequential positions are encoded and associated with each other to form profile feature data.

[0036] In the above processing, to improve the stability and reliability of the candidate morphological feature fragment recognition results, a forward scan is performed along a preset reading direction, and a reverse scan is performed in the opposite direction. Both forward and reverse scans use the joint determination of the aforementioned side profile edge position change, pixel grayscale change, and connected region range to obtain their respective initial candidate morphological feature fragments. To unify boundary semantics, the termination position recorded in the reverse scan under its scanning order is converted to the corresponding starting position in a unified image coordinate system, and the starting position recorded in the reverse scan under its scanning order is converted to the corresponding termination position in a unified image coordinate system. After conversion, the starting position of the forward scan is compared with the corresponding starting position of the reverse scan, and the termination position of the forward scan is compared with the corresponding termination position of the reverse scan. By performing cross-validation on the same side profile structure through bidirectional scanning, the impact of local false detections, boundary drift, or transient noise introduced by unidirectional scanning can be reduced.

[0037] Furthermore, during the above matching process, multiple reverse scan candidate morphological feature segments may exist corresponding to the same forward scan candidate morphological feature segment. For each reverse candidate morphological feature segment, the absolute value of the positional difference between the forward starting position and the corresponding reverse starting position, as well as the absolute value of the positional difference between the forward ending position and the corresponding reverse ending position, are calculated, and their sum is taken as the matching difference. The reverse candidate morphological feature segments with the smallest matching difference are selected as the matched segments, sorted from smallest to largest. If there are two or more candidate objects with the same matching difference, the one with the larger interval overlap ratio and the smaller absolute value of the starting boundary difference is selected as the priority matching object. Preferably, a reverse candidate morphological feature segment is matched with only one forward candidate morphological feature segment to maintain a one-to-one correspondence.

[0038] After completing the above position matching and confirming that the position difference between corresponding positions is within a preset position difference threshold range, the first position interval is determined based on the forward scan candidate morphological feature segments, and the second position interval is determined based on the corresponding start and end positions of the reverse scan candidate morphological feature segments after unified coordinate conversion. The overlap length between the first and second position intervals is calculated. and merge interval length and with As the overlap ratio. When When the preset overlap ratio threshold is reached, the bidirectional scan result is determined to meet the interval consistency condition and the matching segment is retained; when If the preset overlap ratio threshold is not met, the matched segment is deleted or marked as a low-confidence segment. The overlap ratio is used to determine whether to retain the matched segment, rather than directly specifying the final boundary, to avoid creating two different position intervals with the subsequent boundary averaging correction rule. The interval length is uniformly calculated by subtracting the starting coordinate from the ending coordinate, not by the number of pixels including the endpoints.

[0039] After the matched segments satisfy the interval consistency condition, the first signed position difference is further calculated. Second signed position difference ,in and These represent the start and end positions of the candidate fragments in the forward scan, respectively. and These represent the start and end positions of the candidate segments after unified coordinate conversion during reverse scanning, respectively. and The signs are consistent, and and When none of them exceed the preset offset threshold, Determine the corrected starting position, and The corrected termination position is determined; when the output coordinates are required to be integers, they are rounded to integers according to pre-defined rounding rules (rounding down or rounding up). and If the signs are inconsistent or any absolute value exceeds the preset offset threshold, no average correction will be performed. Instead, the overlapping interval boundary or the unidirectional scan boundary with high confidence can be used as the output according to the preset rules.

[0040] In the process of rule-based screening of candidate morphological feature segments, the starting point, ending point, pixel grayscale change starting point, and connected region boundary of each candidate morphological feature segment are determined. Along the preset reading direction, the grayscale transition corresponding to the true side profile structure usually appears before the clear outer edge offset; therefore, the pixel grayscale change starting point should be located before the starting point of the outer edge change. If the starting point of the outer edge change is located before or coincides with the starting point of the pixel grayscale change, the leading relationship is not established, and the corresponding candidate morphological feature segment is deleted. The ending point of the outer edge change should also be within the effective range of the corresponding foreground connected region; if the ending point of the outer edge change exceeds this range, it indicates that the outer edge change has not converged within the effective connected region, and the corresponding candidate morphological feature segment is deleted.

[0041] Based on this, a first position difference from the starting point of pixel grayscale change to the starting point of outer edge position change, and a second position difference from the ending point of outer edge position change to the boundary of the connected region can be calculated separately in the reading direction. The first position difference characterizes the leading effect of grayscale change on outer edge change, and the second position difference characterizes the degree of closure between the terminating position of outer edge change and the boundary of the connected region. Both the first and second position differences are measured in the reading direction under a unified coordinate system. When the first position difference does not reach the preset leading threshold, it indicates that the guiding effect of pixel grayscale change on outer edge change is insufficient, and the candidate morphological feature segment lacks sufficient support in its formation mechanism; therefore, the corresponding candidate morphological feature segment is deleted. When the second position difference does not reach the preset closure threshold, it indicates that the remaining distance between the terminating point of outer edge change and the boundary of the connected region is too small, failing to reflect a stable boundary convergence relationship. In this case, the confidence of the termination position of the corresponding segment is low, so the corresponding candidate morphological feature segment is also deleted. By setting leading and closure thresholds, the rationality of the retained segments in terms of structural formation order and spatial convergence relationship can be further improved.

[0042] To further enhance the completeness of the screening rules, the outer edge change interval can be determined using the starting point and ending point of the outer edge position change as boundaries, and the constraint interval can be determined using the starting point of the pixel grayscale change and the boundary of the connected region as boundaries. The outer edge change interval reflects the spatial range of the actual change in the outer edge contour, while the constraint interval reflects the allowable range defined by the grayscale change starting point to the effective boundary of the connected region. Then, it is determined whether the outer edge change interval is completely contained within the constraint interval. If the outer edge change interval is not completely contained within the constraint interval, it indicates that the outer edge change exceeds the effective range jointly defined by the grayscale change and the connected region, suggesting that the candidate morphological feature segment has boundary overflow, positioning offset, or structural incompleteness issues, and therefore the corresponding candidate morphological feature segment is deleted. Through the joint screening of the starting point sequence relationship, the ending point closure relationship, and the interval inclusion relationship, the retained candidate morphological feature segments can simultaneously meet the correspondence requirements between outer edge change, grayscale change, and connected region constraints, thus providing more accurate basic data for the consistency verification and sequential position encoding generation of subsequent profile feature segments.

[0043] After determining the profile feature segments as described above, the start position, end position, and length of each profile feature segment in the reading direction are further determined. The start position represents the beginning point of the corresponding profile feature segment in the reading direction, the end position represents the end point of the corresponding profile feature segment in the reading direction, and the length is determined based on the range between the start and end positions. After extracting the position parameters of each profile feature segment, they are sorted according to their order of arrangement in the reading direction, and sequential position codes are generated by combining the positional interval relationship and interval relationship between adjacent profile feature segments. The positional interval relationship reflects the distance feature between two adjacent profile feature segments, and the interval relationship reflects the relative state of two adjacent profile feature segments in spatial distribution. Specifically, the interval relationship includes separation, contiguousness, overlap, and inclusion. When there is a significant gap between two adjacent profile feature segments, it can be determined as a separation relationship; when the start position of the latter profile feature segment is greater than the end position of the former profile feature segment, and the gap between them is within a preset separation condition range, it can be recorded as separation. When the boundaries of two adjacent profile feature segments are connected without a significant gap, they can be identified as connected. When the starting position of the latter profile feature segment is the same as the ending position of the former profile feature segment, or the difference between them is within a preset connection tolerance range, they can also be considered connected. When partial intervals of two adjacent profile feature segments overlap, they can be identified as overlapping. When the starting position of the latter profile feature segment is within the interval of the former profile feature segment, and the ending position of the latter profile feature segment exceeds the ending position of the former profile feature segment, they can be considered as overlapping. When the entire interval of one profile feature segment is within the interval of another profile feature segment, they can be identified as containing. By encoding the above spatial relationships between profile feature segments, the relative positional distribution of multiple feature segments in the profile structure can be transformed into a structured representation with a sequence, thereby improving the usability of profile feature data in comparison, retrieval, and subsequent analysis.

[0044] Furthermore, to improve the ability of sequential position coding to represent fine-grained positional differences, hierarchical sequential position coding can be generated based on the positional difference between two adjacent profile feature segments in the reading direction, the ratio of the length of the preceding profile feature segment to the length of the following profile feature segment, and the interval relationship. In specific processing, the basic coding category is first determined according to the interval relationship, then the positional level is determined according to the interval containing the positional difference, and the scale level is determined according to the interval containing the length ratio. Finally, the basic coding category, positional level, and scale level are combined to form a hierarchical sequential position coding. The phrase "interval relationships belonging to the same category" means that different pairs of adjacent profile feature segments are all determined to have the same basic relationship, such as separation, connection, overlap, or inclusion; for example, the first and second adjacent segment pairs are both separation relationships. For these different adjacent segment pairs with the same basic relationship, further hierarchical classification is performed based on their respective positional differences and length ratios, such as near / far distance and low / high proportion, rather than referring to the existence of two interval relationships within the same pair of segments that need to be compared for consistency.

[0045] The following are several specific embodiments of the application of the present invention.

[0046] Example 1 In this embodiment, orthodontic lateral profile images are first acquired using a lateral profile imaging device. These images are then subjected to pose correction, scale normalization, and grayscale normalization sequentially within a unified image coordinate system to obtain standardized images. Pose correction reduces the impact of head pitch, rotation, and slight sway on the positioning of the lateral profile's outer edge. Scale normalization eliminates size differences caused by varying acquisition distances and resolutions. Grayscale normalization reduces interference from lighting fluctuations and device response differences on grayscale transition recognition. After these processes, the spatial positional relationships and grayscale distribution of the lateral profile subject tend to be consistent, thus providing a stable foundation for subsequent scanning and extraction.

[0047] After obtaining the standardized image, the image is scanned along a preset reading direction, and the changes in the outer edge position, pixel grayscale changes, and foreground connected region range are extracted at each scan position. For scan positions that continuously reach the outer edge change threshold, an outer edge change interval is formed. The starting point of pixel grayscale change is determined within a preset association window before the starting point of the outer edge position change, and the projection range of the foreground connected region carrying this outer edge change interval is determined. When the outer edge change interval is within this projection range, the starting point of pixel grayscale change is no later than the starting point of the outer edge position change, and both belong to the same local region, the corresponding local interval is initially determined as a candidate morphological feature segment, and the starting and ending points of the outer edge position change, the starting point of pixel grayscale change, and the boundary of the connected region range are associated and recorded.

[0048] To improve the stability of candidate morphological feature fragment localization results, the standardized image was further scanned in both forward and reverse directions based on the above scanning. Forward scanning was performed along a preset reading direction, while reverse scanning was performed in the opposite direction. Since the two scanning methods read the same lateral structure in opposite order, the boundary representation of the same morphological fragment in forward and reverse scanning has a mirror-image correspondence. Cross-validation of the same region through bidirectional scanning can reduce the impact of local noise, instantaneous grayscale fluctuations, and boundary drift on the candidate fragment extraction results under a single scanning direction.

[0049] In the candidate fragment matching stage, each candidate morphological feature fragment obtained from the reverse scan is first converted to a coordinate expression consistent with that of the forward scan. Then, for each candidate morphological feature fragment obtained from the forward scan, the sum of the absolute values ​​of the corresponding positional differences between it and each candidate morphological feature fragment obtained from the reverse scan is calculated, and this value is used as the matching optimization index. The matching calculation formula used is:

[0050] In the formula, This represents the matching difference between the j-th reverse candidate morphological feature segment and the current positive candidate morphological feature segment. This indicates the starting position of the positive candidate morphological feature segment. This indicates the termination position of a positive candidate morphological feature segment. and These represent the start and end positions of the j-th reverse candidate morphological feature segment after unified coordinate transformation, respectively. The superscript F indicates forward scanning, and the superscript R indicates the coordinates of the reverse scanning result after unified coordinate transformation. The end position recorded in the reverse scanning sequence is converted to... The starting position recorded in the reverse scan sequence is converted to... By summing the absolute values ​​of the deviations between two sets of corresponding boundaries, the consistency between the two boundary points can be reflected simultaneously; according to Sort by size from smallest to largest, and take... The smallest one is used as the matching object.

[0051] In the example data of this embodiment, the forward scan obtains a candidate morphological feature segment with a starting position of 118 pixels and an ending position of 146 pixels. The reverse scan obtains two candidate objects, where the first reverse candidate segment corresponds to positions from 117 pixels to 145 pixels, and the second reverse candidate segment corresponds to positions from 121 pixels to 150 pixels. Substituting the above data into the aforementioned calculation formula, the matching difference of the first reverse candidate segment can be obtained as: The matching difference of the second reverse candidate segment is Meanwhile, setting the corresponding position difference threshold to 5 pixels, the two sets of boundary position differences corresponding to the first reverse candidate segment are 1 pixel and 1 pixel respectively, and the two sets of boundary position differences corresponding to the second reverse candidate segment are 3 pixels and 4 pixels respectively. Although neither exceeds the position difference threshold, the matching difference of the first reverse candidate segment is smaller. Therefore, the first reverse candidate segment is determined as the unique matching segment, while the second reverse candidate segment is discarded. (Appendix) Figure 2 The matching process is illustrated, where the forward candidate segment interval, the reverse candidate segment 1 interval, the reverse candidate segment 2 interval, and the bidirectional boundary correspondence are all represented on a unified reading direction coordinate axis. (See attached diagram) Figure 2 It can be intuitively seen that candidate object 1 is closer to the positive candidate fragment at both the start boundary and the end boundary than candidate object 2. Therefore, when multiple reverse candidate objects coexist, the matching result with the closer overall boundary can be retained first, thus improving the consistency of candidate fragment localization.

[0052] After identifying a unique matching segment, the first position interval is determined based on the forward candidate morphological feature segment, and the second position interval is determined based on the matched reverse candidate morphological feature segment. The overlap ratio between the two is then calculated to determine whether the bidirectional scanning results have sufficient consistency. The formula used to calculate the overlap ratio is:

[0053] In the formula, This indicates the overlap ratio between the first position interval and the second position interval. Indicates the length of the overlapping portion of the two intervals. This represents the overall coverage length after merging the two intervals. The above calculation formula is based on the fact that if the bidirectional scans locate the same candidate segment consistently, the common coverage portion of the two intervals should account for a high proportion of the overall coverage area; if the two intervals only have a small overlap, it indicates that the bidirectional scans have significant disagreements in determining the boundary of the segment, and it should not be directly retained. Therefore, in this embodiment, the overlap ratio is used as a consistency criterion for whether to retain a candidate segment, rather than being directly used as the final boundary output result.

[0054] In this embodiment, the length of each interval is calculated based on the difference between the ending coordinate and the starting coordinate. For example, the length of the interval from 118 pixels to 145 pixels is recorded as 145-118=27 pixels, instead of being calculated based on the number of pixels including both endpoints.

[0055] Using the example data above, the first position range of the positive candidate morphological feature segment is 118 pixels to 146 pixels, and the second position range of the first negative candidate segment is 117 pixels to 145 pixels. The overlap between the two is 118 pixels to 145 pixels, therefore... The value is 27 pixels, and the combined coverage area of ​​the two is from 117 pixels to 146 pixels. The overlap ratio is 29 pixels, resulting in R = 27 / 29, which is approximately 93.1%. When the overlap ratio threshold is set to 80%, this candidate segment meets the retention criteria, indicating that bidirectional scanning has high consistency with its localization results. If the aforementioned second reverse candidate segment is used as the comparison object, its overlap length with the forward candidate segment is 25 pixels, the merge length is 32 pixels, and the resulting overlap ratio is approximately 78.1%, which is lower than 80%, indicating insufficient interval-level consistency and therefore should not be retained. (Appendix) Figure 3 The above determination process is illustrated, where the ranges of the forward candidate segments, reverse candidate segment 1, and reverse candidate segment 2, the overlapping portion of the two intervals, and the reference line for the overlap ratio determination threshold are all given in the same coordinate system. (See attached diagram.) Figure 3 It can be further seen that the overlapping area of ​​segment 1 accounts for a significantly higher proportion of the overall coverage than that of segment 2. Therefore, after completing the boundary matching, it is possible to further eliminate candidate objects with large interval deviations, making the segments that enter the subsequent boundary correction more reliable.

[0056] After the candidate fragments pass the overlap ratio consistency determination, the first signed position difference between the starting boundary of the forward candidate fragment and the corresponding starting boundary of the reverse candidate fragment after unified coordinate transformation is further calculated. And the second signed position difference between the termination boundary of the positive candidate segment and the corresponding termination boundary of the negative candidate segment after unified coordinate transformation. .in, , .like and If the signs are consistent and the absolute values ​​do not exceed the preset offset threshold, then follow the... and Calculate the corrected start and end positions respectively; if and If the signs are inconsistent, or if any absolute value exceeds the preset offset threshold, no averaging correction will be performed. Therefore, the overlap ratio is responsible for determining whether the matched segments meet the interval consistency requirement, while the boundary averaging is responsible for providing a unique correction boundary in the case of a small offset in the same direction.

[0057] Taking the aforementioned retained segment as an example, the first signed position difference is 118-117=1, and the second signed position difference is 146-145=1. Both are positive, and when the offset threshold is set to 3 pixels, their absolute values ​​do not exceed this threshold. Therefore, it can be determined that the candidate segment has a small amplitude offset in the same direction. In this case, averaging over the starting boundary yields 117.5 pixels, and averaging over the ending boundary yields 145.5 pixels. When the output boundary is represented using integer pixels, the final starting position can be determined as 118 pixels and the final ending position as 146 pixels according to a unified rounding rule. (Appendix) Figure 4 The correction process is illustrated, where the reverse scan boundary, forward scan boundary, average correction boundary, and final output interval are all represented on a unified read direction coordinate axis, and the start and end boundaries are shown converging towards the average correction boundary. (See attached diagram.) Figure 4 It can be more clearly seen that both the forward and reverse corresponding boundaries exhibit small, consistent offsets at both ends. After averaging correction, the boundary positions converge towards the center and are ultimately projected into a unified output boundary range. This averaging correction process reduces the systematic bias introduced by a single scanning direction while maintaining the overall stability of the segment's range, resulting in more balanced and accurate boundary positioning.

[0058] After the above standardization process, candidate fragment extraction, bidirectional scanning matching, overlap ratio consistency determination, and same-direction offset boundary correction, the obtained candidate morphological feature fragments have high spatial consistency and boundary reliability, and can be used as the basis data for subsequent rule filtering and structured output. This embodiment shows that by first using the sum of the absolute values ​​of the corresponding position differences to achieve a unique match among multiple candidate objects, then using the overlap ratio to determine whether to retain the candidate fragment, and using average correction to determine the final boundary when the conditions are met, the influence of local noise, boundary drift, and scanning direction differences on the positioning results can be effectively reduced, thereby improving the positioning stability and boundary accuracy of candidate morphological feature fragments.

[0059] Example 2 In this embodiment, an orthodontic profile image is first acquired, and then, under a unified image coordinate system, pose correction, scale normalization, and grayscale normalization are performed sequentially to obtain a standardized image. Pose correction reduces the impact of head pitch, rotation, and slight sway on the positioning of the profile's outer edge; scale normalization eliminates size differences caused by different acquisition distances and resolutions; and grayscale normalization reduces interference from illumination fluctuations and imaging response differences on grayscale transition recognition. After standardization, the profile subject has a unified basis in spatial position and grayscale distribution, thus facilitating stable scanning along a preset reading direction.

[0060] After obtaining the standardized image, it is scanned step by step along a preset reading direction, and the changes in the outer edge position, pixel grayscale changes, and foreground connected region range are extracted at each scanning position. For local intervals where the outer edge position change and grayscale change correspond within the same region, and this region is located within a reasonable foreground connected region, it is initially identified as a candidate morphological feature segment. For each candidate morphological feature segment, its corresponding pixel grayscale change start point, outer edge position change start point, outer edge position change end point, and connected region range boundary are further determined. Specifically, the pixel grayscale change start point reflects the position where a stable grayscale state transitions to a distinct transition state; the outer edge position change start point reflects the position where the contour begins to show a identifiable shift or turn; the outer edge position change end point reflects the position where the contour change tends to end; and the connected region range boundary reflects the effective termination boundary of the foreground connected region carrying this local structure in the reading direction.

[0061] In the rule-based screening process for candidate morphological feature fragments, the process first determines whether a leading relationship is satisfied between the starting point of pixel grayscale change and the starting point of outer edge position change. Then, it determines whether a closure relationship is satisfied between the ending point of outer edge position change and the boundary of the connected region. The leading relationship is required because in the reading direction, the true profile structure typically shows a grayscale transition first, followed by a clear contour shift. If the outer edge change occurs before the grayscale change, or the interval between the two is too small, the fragment is more likely caused by local noise, shadow boundaries, or false detections. The closure relationship is required because the true contour change should converge within the effective foreground connected region. If the ending point of the outer edge change approaches or crosses the boundary of the connected region, it indicates that the fragment may be affected by background interference or boundary drift, making it difficult to output as a stable feature.

[0062] To provide a quantifiable basis for determining the leader and closure relationships, this embodiment defines the leader position difference and closure position difference as follows:

[0063] In the formula, This represents the leading position difference from the starting point of the pixel grayscale change to the starting point of the outer edge position change. This represents the difference in closed position from the endpoint of the change in position at the outer edge to the boundary of the connected region. Indicates the starting point of pixel grayscale change. Indicates the starting point of the change in the outer edge position. Indicates the endpoint of the change in the outer edge position. Indicates the boundary of a connected region. The minus sign is used to indicate the coordinate difference along the reading direction. It can directly reflect the degree to which grayscale changes lead to contour changes. It can directly reflect whether there is still sufficient convergence space between the termination of contour changes and the effective boundary. When the difference in the leading position is large, it indicates that the grayscale transition has a sufficient guiding effect on contour formation; when the difference in the closing position is large, it indicates that the contour termination position is located inside the connected region, and the boundary stability is good.

[0064] Taking a candidate morphological feature segment as an example, its pixel grayscale change starts at 95 pixels, the outer edge change starts at 101 pixels, the outer edge change ends at 138 pixels, and the boundary of the connected region is 144 pixels. Substituting the above data into the aforementioned calculation formula, we can obtain... =101-95=6 pixels =144-138=6 pixels. If the leading threshold is set to 4 pixels and the closing threshold is set to 3 pixels, then the candidate morphological feature fragment simultaneously satisfies... Not less than 4 pixels and The condition of being no less than 3 pixels is therefore retained. Conversely, if the starting point of the pixel grayscale change in another candidate segment is 99 pixels, the starting point of the outer edge position change is still 101 pixels, the ending point of the outer edge position change is 143 pixels, and the boundary of the connected region is 144 pixels, then... =2 pixels, =1 pixel, neither of which reached the corresponding threshold, indicating that the segment is unstable in terms of formation order and boundary convergence, and should be deleted. (Attached) Figure 5 The quantitative screening process is illustrated below. (Attached) Figure 5 The horizontal axis sequentially identifies four key locations: the starting point of grayscale change, the starting point of outer edge change, the ending point of outer edge change, and the boundary of the connected region. Two broken lines represent the positional changes of candidate segments to be retained and those to be deleted. Candidate segments to be retained exhibit large effective intervals in both the leading and closing position differences, while candidate segments to be deleted show insufficient intervals in both. This quantitative screening effectively eliminates unstable candidate segments caused by local noise and boundary drift.

[0065] After determining the leading and closing relationships, the outer edge change range is further defined by the starting and ending points of the outer edge position change, and the constraint range is defined by the starting point of the pixel grayscale change and the boundary of the connected region. Then, it is determined whether the outer edge change range is completely contained within the constraint range. Taking the previously retained segment as an example, its outer edge change range is from 101 pixels to 138 pixels, and the constraint range is from 95 pixels to 144 pixels. The former completely falls within the latter, indicating that the entire contour change process of this segment is within the effective range jointly defined by the grayscale change and the foreground connected region, thus exhibiting good structural integrity. Furthermore, taking another pair of proportional candidate segments as an example, its pixel grayscale change starting point is 100 pixels, the outer edge position change starting point is 103 pixels, the outer edge position change ending point is 147 pixels, and the connected region boundary is 144 pixels. Therefore, its outer edge change range is from 103 pixels to 147 pixels, and the constraint range is from 100 pixels to 144 pixels. Because the ending position of the outer edge variation range at 147 pixels exceeds the ending position of the constraint range at 144 pixels, this outer edge variation range is not completely contained within the constraint range. This indicates that the segment has issues such as boundary overflow, positioning offset, or structural incompleteness, and should be deleted. (Attached) Figure 6 The inclusion relationship of this interval is illustrated. (See attached diagram) Figure 6 The constraint interval and the outer edge variation interval are represented by strip-shaped intervals. For the retained segments, the outer edge variation interval falls completely inside the constraint interval; for the comparative segments that fail the screening, the end of their outer edge variation interval exceeds the boundary of the constraint interval, thus forming a clear overflow state. By combining the constraints of leading relations, closure relations, and interval inclusion relations, the final retained segments can better conform to the formation rules of the real side profile structure.

[0066] In adjacent scan consistency screening, for example, if the candidate segments obtained for the same local structure at three consecutive scan positions are [101, 138], [102, 139], and [100, 137], then the starting position range is... =102-100=2 pixels, ending position range =139-137=2 pixels, all three segments are 37 pixels long, and the length has a relative range. =0. When the preset start position threshold and end position threshold are both 3 pixels, and the length relative range threshold is 15%, the group of candidate segments meets the adjacent scan consistency requirement. Conversely, if one of the segments becomes [106, 145], the start position range and end position range reach 6 pixels and 8 pixels respectively, exceeding the thresholds. It can be determined that the boundary change is too large, and the unstable candidate segment or the entire group of candidate results is deleted. The above values ​​are only non-limiting examples to illustrate the judgment method. The actual thresholds can be set according to the standardization scale and sample statistical results.

[0067] For the retained segments after screening, their start position, end position, and length in the reading direction are further determined. Extracting the start and end positions clarifies the boundary range of each profile feature segment within the overall profile structure; extracting the length allows for comparison of scale differences and organizational morphology among different segments. After extracting the positional parameters, the profile feature segments are first sorted according to their order in the reading direction, and then the positional interval and interval relationships between adjacent segments are determined. When the start position of a later segment is greater than the end position of a previous segment, it is considered a separation relationship; when their boundaries are adjacent and the interval is close to zero, it is considered a contiguous relationship; when some intervals of the two segments overlap, it is considered an overlapping relationship; when one interval completely falls within the other interval, it is considered an inclusion relationship. Taking two adjacent retained segments as an example, the first segment's interval is from 101 to 138 pixels, and the second segment's interval is from 145 to 162 pixels. After sorting according to the reading direction, the first segment is located in front, and the second segment is located behind, with a segment interval distance of 7 pixels. Therefore, their interval relationship can be determined as separation. Based on this result, a sequential position code containing the order of segments and the identifier of separation relationship can be generated to reflect the arrangement relationship of two segments in the reading direction.

[0068] During encoding output, a sequence identifier S1 can be assigned to the previous segment, a sequence identifier S2 can be assigned to the next segment, and a relationship identifier SEP can be assigned to adjacent segment pairs consisting of S1 and S2. Thus, either "sequence identifier + relationship identifier with the next segment" can be appended to the previous segment as a field, or {S1, S2, SEP} can be used as an independent encoding record for adjacent segment pairs; both forms are consistent with the limitation in the claim of "generating sequential position codes according to the positional interval relationship and interval relationship between adjacent profile feature segments".

[0069] To further enhance the encoding's ability to represent fine-grained structural differences, this embodiment introduces the ratio of adjacent segment lengths as a grading criterion, building upon sequential position encoding. The formula for calculating the length ratio is as follows:

[0070] In the formula, This represents the ratio of the length of the preceding facial feature segment to the length of the following facial feature segment. Indicates the length of the preceding facial feature segment. Indicates the length of the latter facial feature segment. and These represent the start and end positions of the preceding facial feature segment, respectively. and These represent the starting and ending positions of the subsequent facial feature segment, respectively. Simple interval relationships can only reflect whether adjacent segments are separated, connected, overlapping, or contained, but cannot distinguish the structural differences corresponding to different scale combinations when they belong to the same separation relationship; after introducing the length ratio, it is possible to further distinguish different organizational states such as the former segment being significantly longer than the latter segment, the two segments being close in length, or the latter segment being significantly longer than the former segment.

[0071] In the example data of this embodiment, the first facial feature segment extends from 101 pixels to 138 pixels, with a length of 37 pixels, and the second facial feature segment extends from 145 pixels to 162 pixels, with a length of 17 pixels. Therefore, =37 / 17, its value is approximately 2.18. Meanwhile, the segment interval between the two is 7 pixels, indicating a separation relationship. The segment interval can be further divided into near-distance and far-distance levels according to preset intervals, and the length ratio can be divided into low-ratio and high-ratio levels according to preset intervals. For example, 0 pixels to 10 pixels are classified as near-distance, and greater than 10 pixels as far-distance; 1 to 1.5 is classified as low-ratio, and greater than 1.5 as high-ratio. Thus, the two adjacent segments can be categorized into separation, near-distance, and high-ratio levels, and a hierarchical sequential position code can be generated accordingly. This hierarchical sequential position code not only indicates that the two segments are separated in the reading direction, but also further indicates that they belong to a combination state with a small spacing and a significant length difference, thereby enabling a more refined expression of the local organizational features of the lateral structure.

[0072] After the above processing, this embodiment first filters candidate morphological feature segments layer by layer through the preceding relationship, the closing relationship and the interval inclusion relationship, then determines the final side profile feature segment by extracting the start position, the end position and the length, and generates sequential position code and hierarchical sequential position code on this basis.

[0073] Accordingly, embodiments of the present invention also provide a computer vision-based orthodontic lateral profile image feature extraction system, such as... Figure 7 The image shown is a schematic diagram of the system. This computer vision-based orthodontic lateral profile image feature extraction system includes the following modules: Image preprocessing unit 701 acquires orthodontic profile image and performs standardized preprocessing on the orthodontic profile image to obtain standardized image; The candidate morphological feature segment determination unit 702 scans the standardized image along a preset reading direction and determines candidate morphological feature segments based on changes in the outer edge position of the side profile, changes in pixel grayscale, and the range of connected regions. The candidate morphological feature segment filtering unit 703 performs rule-based filtering on the candidate morphological feature segments and deletes candidate morphological feature segments that do not meet the rules. The feature segment verification unit 704 verifies the candidate morphological feature segments retained after screening according to the consistency of the start position, end position and length, and determines the side profile feature segments. The position encoding unit 705 generates sequential position codes based on the arrangement order and interval relationship of the side profile feature fragments in the reading direction, and outputs the side profile feature data.

[0074] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0075] The present invention also provides a storage medium, which is a computer-readable storage medium storing a computer program thereon, the computer program being executable when it runs. Figure 1 The method shown may include some or all of the steps. The storage medium may include read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc. The storage medium may also include non-volatile memory or non-transitory memory, etc.

[0076] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data provider to another website, computer, server, or data provider via wired or wireless means.

[0077] The embodiments of the present invention have been described in detail above. Specific implementation methods have been used to illustrate the present invention. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and systems of the present invention, and are merely some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention, and the content of this specification should not be construed as a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for extracting features from orthodontic lateral profile images based on computer vision, characterized in that, The method includes: Step 1: Obtain orthodontic profile images and perform standardized preprocessing on the orthodontic profile images to obtain standardized images; Step 2: Scan the standardized image along the preset reading direction, and determine candidate morphological feature segments based on changes in the outer edge position of the profile, changes in pixel grayscale, and the range of connected regions. Step 3: Perform rule-based filtering on candidate morphological feature segments and delete candidate morphological feature segments that do not meet the rules; Step 4: Verify the consistency of the candidate morphological feature fragments retained after screening according to the start position, end position and length to determine the side profile feature fragments. Step 5: Generate sequential position codes based on the arrangement order and interval relationship of the profile feature fragments in the preset reading direction, and output the profile feature data.

2. The method for extracting orthodontic lateral profile image features based on computer vision according to claim 1, characterized in that, In step 2, the preset reading direction is preset according to the image coordinate system or the unfolding direction of the side profile. Based on the joint judgment of three dimensions, namely the change of the outer edge position of the side profile, the change of pixel gray level, and the range of connected regions, multiple candidate morphological feature segments are determined. The candidate morphological feature segments are local intervals with start and end positions in the preset reading direction. The local intervals correspond to the morphological changes in the side profile that may have diagnostic or structural significance.

3. The method for extracting orthodontic lateral profile image features based on computer vision according to claim 2, characterized in that, In step 2, during scanning, a forward scan is performed along the preset reading direction, and a reverse scan is performed in the opposite direction to the preset reading direction. In both the forward and reverse scans, candidate morphological feature fragments are obtained based on the joint determination of changes in the side profile outer edge position, changes in pixel grayscale, and the range of connected regions. The start and end positions of the candidate morphological feature segments obtained by forward scanning are matched with the end and start positions of the candidate morphological feature segments obtained by reverse scanning, respectively. When the position difference between the corresponding positions is within the preset position difference threshold range, the match is successful, and the successfully matched segments are retained as candidate morphological feature segments. The matching process includes: when there are multiple candidate morphological feature segments obtained by reverse scanning corresponding to the candidate morphological feature segments obtained by the same forward scan, calculating the sum of the absolute values ​​of the position differences between the corresponding positions of the candidate morphological feature segments obtained by each reverse scan, sorting them in ascending order of the sum of the absolute values ​​of the position differences, and determining the candidate morphological feature segment obtained by reverse scanning with the smallest sum of the absolute values ​​of the position differences as the segment to be matched with the candidate morphological feature segment obtained by the forward scan.

4. The method for extracting orthodontic lateral profile image features based on computer vision according to claim 3, characterized in that, In step 2, after the position difference between corresponding positions is within the preset position difference threshold range, the first position interval is determined based on the candidate morphological feature fragments obtained by forward scanning, and the second position interval is determined based on the candidate morphological feature fragments obtained by reverse scanning; when the ratio of the overlap interval length to the merged interval length of the first position interval and the second position interval reaches the preset overlap ratio threshold, it is determined that the candidate morphological feature fragments obtained by forward scanning and the candidate morphological feature fragments obtained by reverse scanning satisfy the interval consistency condition. When the interval consistency condition is met, a first signed position difference is calculated between the starting position of the candidate morphological feature segment obtained by forward scanning and the corresponding starting position of the candidate morphological feature segment obtained by reverse scanning after unified coordinate conversion, and a second signed position difference is calculated between the ending position of the candidate morphological feature segment obtained by forward scanning and the corresponding ending position of the candidate morphological feature segment obtained by reverse scanning after unified coordinate conversion. When the signs of the first signed position difference and the second signed position difference are consistent and their absolute values ​​are within a preset offset threshold range, the arithmetic mean of the starting position obtained by forward scanning and the corresponding starting position after unified coordinate conversion is determined as the starting position of the candidate morphological feature segment, and the arithmetic mean of the ending position obtained by forward scanning and the corresponding ending position after unified coordinate conversion is determined as the ending position of the candidate morphological feature segment. The corresponding starting position after unified coordinate conversion and the corresponding ending position after unified coordinate conversion are respectively obtained by converting the ending position and the starting position recorded by reverse scanning in the scanning order.

5. The method for extracting orthodontic lateral profile image features based on computer vision according to claim 1, characterized in that, In step 3, candidate morphological feature segments are filtered according to rules based on the continuity of the candidate morphological feature segments in the preset reading direction, the positional correspondence between pixel grayscale changes and outer edge position changes, and the consistency of adjacent scan results. Specifically, this includes: Step 31: Determine the starting point of the outer edge position change, the ending point of the outer edge position change, the starting point of the pixel grayscale change, and the boundary of the connected region for each candidate morphological feature segment. Step 32: Determine the sequential relationship between pixel grayscale change and outer edge position change, as well as the closure relationship of the outer edge position change region. When the starting point of the outer edge position change is located before or coincides with the starting point of the pixel grayscale change in the preset reading direction, or when the ending point of the outer edge position change is not located within the connected region, delete the corresponding candidate morphological feature fragment. Step 33: Determine the positional correspondence between pixel grayscale changes and outer edge position changes. Calculate a first positional difference from the starting point of the pixel grayscale change to the starting point of the outer edge position change, and a second positional difference from the ending point of the outer edge position change to the boundary of the connected region, respectively, along the reading direction. When the first positional difference does not reach a preset leading threshold, or the second positional difference does not reach a preset closing threshold, delete the corresponding candidate morphological feature segment. Determine the outer edge change interval using the starting point and ending point of the outer edge position change as boundaries, and determine the constraint interval using the starting point of the pixel grayscale change and the boundary of the connected region, respectively. When the outer edge change interval is not completely contained by the constraint interval, delete the corresponding candidate morphological feature segment. Step 34: Filter based on the consistency of adjacent scan results. If a candidate morphological feature segment is detected only at a single scan position and cannot be supported in adjacent scan positions, or if the range of the start position, the range of the end position, or the relative range of the length of the same local structure in multiple adjacent scan results exceeds the corresponding preset fluctuation threshold, then the candidate morphological feature segment is deemed to have insufficient stability and is deleted.

6. The method for extracting orthodontic lateral profile image features based on computer vision according to claim 1, characterized in that, In step 4, when the starting position of the candidate morphological feature fragment retained after screening remains stable within the allowable fluctuation range, the ending position remains stable within the allowable fluctuation range, and the length of the candidate morphological feature fragment does not show abnormal expansion or contraction, the candidate morphological feature fragment is determined as a valid profile feature fragment. The length of the candidate morphological feature fragment is determined by the range between the ending position and the starting position of the candidate morphological feature fragment. The allowable fluctuation range is preset according to the image resolution, standardization scale, or sample statistical results.

7. The method for extracting orthodontic lateral profile image features based on computer vision according to claim 1, characterized in that, In step 5, the start position, end position, and length of each profile feature segment in the preset reading direction are determined; sequential position codes are generated according to the positional interval relationship and interval relationship between adjacent profile feature segments; wherein, the interval relationship includes separation, connection, overlap, and inclusion.

8. The method for extracting orthodontic lateral profile image features based on computer vision according to claim 7, characterized in that, In step 5, a hierarchical sequential position code is generated based on the position difference between two adjacent profile feature segments in the preset reading direction, the ratio of the length of the preceding profile feature segment to the length of the following profile feature segment, and the interval relationship; for different adjacent profile feature segment pairs whose interval relationship belongs to the same category, the hierarchical sequential position code is further divided based on the position difference and the length ratio.

9. A computer vision-based orthodontic lateral profile image feature extraction system, characterized in that, The system includes: The image preprocessing unit acquires orthodontic profile images and performs standardized preprocessing on the orthodontic profile images to obtain standardized images. The candidate morphological feature segment determination unit scans the standardized image along a preset reading direction and determines candidate morphological feature segments based on changes in the outer edge position of the profile, changes in pixel grayscale, and the range of connected regions. The candidate morphological feature segment filtering unit performs rule-based filtering on candidate morphological feature segments and deletes candidate morphological feature segments that do not meet the rules. The feature fragment verification unit verifies the candidate morphological feature fragments retained after screening according to the consistency of the start position, end position and length, and determines the profile feature fragments. The position encoding unit generates sequential position codes based on the arrangement order and interval relationship of the profile feature fragments in the reading direction, and outputs the profile feature data.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it performs the steps of the orthodontic lateral profile image feature extraction method based on computer vision as described in any one of claims 1 to 8.

Images