An adaptive palmprint multidimensional quantitative analysis method and system integrating anatomical features

Abstract

This invention provides an adaptive palmprint multi-dimensional quantitative analysis method and system that integrates anatomical features. It offers two technical paths: anatomical guidance and deep learning segmentation. This approach leverages prior medical knowledge for precise derivation and utilizes deep learning models for powerful recognition. Both methods can be used independently or in combination, broadening the applicability of the technology. By dynamically linking key parameters of the Canny and Hough algorithms to image resolution and ROI size, it ensures detection stability under images of different qualities and sizes. The use of dynamically generated palmprint ROIs from anatomical reference points to validate and constrain the output of the palmprint segmentation model effectively solves errors that may occur in AI models that do not conform to physiological common sense, making the recognition results more reliable and accurate. Through techniques such as skeletonization and topology analysis, it achieves precise quantification of palmprint features such as thickness, curvature, and branching morphology, which were previously impossible to objectively measure, providing unprecedented data dimensions for the scientific research of hand diagnosis.

Description

Technical Field

[0001] This invention relates to the field of palmprint analysis technology, and in particular to an adaptive palmprint multidimensional quantitative analysis method and system that integrates anatomical features. Background Technology

[0002] Hand diagnosis, as an important part of traditional medicine, assesses a person's health by observing hand features such as palm lines and complexion, with palm line analysis being the core component. Traditional palm line analysis relies entirely on manual observation and judgment by experienced physicians. This method has the following drawbacks: 1) High subjectivity: Different physicians have different diagnostic standards and experiences, leading to a lack of consistency and repeatability in the results; 2) Low efficiency: It cannot be used for large-scale, rapid screening; 3) Qualitative rather than quantitative: It is mostly based on vague descriptions such as "deep / shallow" or "long / short," lacking precise quantitative indicators.

[0003] With the development of computer vision technology, some automated palmprint recognition methods have emerged. However, most existing technologies focus on palmprint identification rather than health analysis, and the few attempts used for health analysis also have shortcomings.

[0004] 1. Lack of adaptability and use of fixed templates: Most methods use fixed region of interest (ROI) templates to find palm lines, which cannot adapt to the differences in individual palm size, shape and posture, resulting in inaccurate localization and low recognition rate in practical applications.

[0005] 2. Ignoring key anatomical information: Existing technologies rarely utilize the inherent and stable anatomical features of the palm (such as the indentation between the fingers) as a positioning benchmark, resulting in a lack of scientific basis for ROI division. This is especially true for secondary lines with variable locations, which are almost impossible to detect effectively.

[0006] 3. Shallow analytical dimensions and weak quantification capabilities: Most techniques are limited to detecting the presence or absence of palm lines or simple length estimation, failing to perform multi-dimensional and precise quantification of key features reflecting pathological conditions, such as the accurate curvature, thickness variation, number of branches and angles of the lines.

[0007] 4. Limitations of deep learning applications: Some deep learning solutions typically perform simple object detection or coarse segmentation, but lack integration with anatomical knowledge, which may lead to errors that do not conform to physiological common sense (such as identifying the heart line as the wrist), resulting in insufficient robustness. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides a method and system for evaluating the modification of traditional Chinese medicine prescriptions based on a quantum-inspired Hamiltonian model. This method transforms the abstract theory of traditional Chinese medicine compatibility into quantifiable, comparable, and interpretable evaluation results through vector representation, Hamiltonian modeling, and energy calculation, providing strong technical support for the precise modification of traditional Chinese medicine prescriptions.

[0009] In a first aspect, the present invention provides an adaptive palmprint multidimensional quantitative analysis method that integrates anatomical features, comprising the following steps:

[0010] The process involves acquiring a palm image, converting the palm image from the RGB color space to the YCrCb color space, generating a preliminary palm binary mask based on a preset skin color range threshold, and using morphological operations with a large kernel to remove noise to obtain the final palm mask.

[0011] Extract the outer contour of the final palm mask, select the contour with the largest area as the main contour of the palm, calculate the convex defects between the main contour of the palm and its convex hull, sort the convex defects by depth, and select the deepest points as the anatomical reference points of the palm.

[0012] The analysis includes center of gravity position analysis, finger distance pattern analysis, and contour asymmetry analysis of the main contour of the palm. The analysis results are then weighted and scored to identify whether the palm is left or right.

[0013] Based on the main outline of the palm, the position information of the palm anatomical reference points, and the information of the left and right hands, the set of palm ROIs {ROI_1, ..., ROI_14} is determined, and palm line detection and filtering are performed in each palm ROI_i, where ROI_i represents the ROI of the i-th palm line;

[0014] A deep learning semantic segmentation network is pre-built and trained as a palmprint segmentation model. The trained palmprint segmentation model is used to detect palmprints in palm images or sets of palmprint ROIs, resulting in a set of binary masks {Mask_1, ..., Mask_14} that is the same size as the original image and contains 14 channels. Here, Mask_i represents the segmentation mask of the i-th palmprint line.

[0015] The palmprint ROI_i is used to perform cross-validation and result fusion on the segmentation mask Mask_i output by the palmprint segmentation model to obtain a high-confidence palmprint mask Mask_k;

[0016] Multi-dimensional quantization analysis was performed on the high-confidence palmprint mask Mask_k to obtain the palmprint quantization results.

[0017] Optionally, the step of determining the palmprint ROI set {ROI_1, ..., ROI_14} based on the main contour of the palm, the position information of the palm anatomical reference points, and the information of the left and right hands, and performing palmprint line detection and filtering within each palmprint ROI_i, includes:

[0018] Based on the main outline of the palm, the position information of the palm anatomical reference points, and the information of the left and right hands, determine the horizontal starting boundary, horizontal ending boundary, vertical upper boundary, and vertical lower boundary of each palm line, and determine the ROI of each palm line based on the boundaries of each palm line, thus obtaining the palm line ROI set {ROI_1, ..., ROI_14}.

[0019] Preprocessing of palm print ROI_i includes: using Gaussian filtering for noise reduction, using the CLAHE algorithm to enhance contrast, and using directional morphology kernels to enhance palm texture;

[0020] The Canny algorithm is used to perform edge detection on the preprocessed palmprint ROI_i, and the Hough algorithm is used to extract line segments. The parameters of the Canny and Hough algorithms are dynamically adjusted according to the resolution of the palm image and the size of the palmprint ROI_i.

[0021] All extracted line segments are initially filtered using preset geometric constraints. A preset quality scoring system is used to comprehensively evaluate the angle deviation, length ratio, and position rationality of each candidate line segment. The line segment with the highest score is selected as the palmprint recognition result.

[0022] Optionally, the step of cross-validating and fusing the segmentation mask Mask_i output by the palmprint segmentation model using palmprint ROI_i includes:

[0023] Calculate the intersection-union ratio (IU / R) of the palmprint ROI_i and the corresponding segmentation mask Mask_i, and compare the IU / R with a preset threshold.

[0024] If the crossover ratio is higher than the preset threshold, the palmprint detection result is confirmed to be a match;

[0025] If the crossover ratio is lower than the preset threshold, it is judged as a false positive and rejected.

[0026] If the traditional method fails to detect the palmprint ROI_i but the palmprint segmentation model does, then the output of the palmprint segmentation model will be adopted as a supplement.

[0027] Optionally, the multi-dimensional quantitative analysis of the high-confidence palmprint mask Mask_k includes:

[0028] The high-confidence palmprint mask Mask_k is processed using a skeletonization algorithm to obtain a skeleton S_k with a width of several single pixels.

[0029] The total number of pixels in the skeleton S_k is calculated to obtain the palmprint length. The ratio of the palmprint length to the effective width or height of the palm is calculated to obtain the standardized palmprint length percentage.

[0030] The average and maximum curvature of the palm print are calculated by analyzing the changes in the tangent direction of the skeleton S_k.

[0031] Iterate through each point p on the skeleton S_k, calculate the local tangent direction of the skeleton at point p, and measure the palm print width on the original mask Mask_p along its normal direction.

[0032] The depth of the palm print is quantified by analyzing the grayscale contrast between the palm print area and the adjacent non-palm print area.

[0033] Topological analysis is performed on the skeleton S_k to automatically identify palm print branch points and endpoints. Based on the palm print branch points and endpoints, the total number of palm print branches, the length of each branch, the angle between the trunk and the branches, and the branch density are calculated.

[0034] In a second aspect, the present invention provides an adaptive palmprint multi-dimensional quantitative analysis system that integrates anatomical features, for implementing the method described in any of the first aspects. The system includes an image acquisition and preprocessing module, an anatomical benchmark extraction module, a palmprint recognition module, and a multi-dimensional feature quantification module.

[0035] The image acquisition and preprocessing module is used to acquire a palm image, convert the palm image from the RGB color space to the YCrCb color space, generate a preliminary palm binary mask based on a preset skin color range threshold, and remove noise using morphological operations with a large kernel to obtain the final palm mask.

[0036] The anatomical reference extraction module is used to extract the outer contour of the final palm mask, select the contour with the largest area as the main contour of the palm, calculate the convex defects between the main contour of the palm and its convex hull, sort the convex defects by depth, and select the deepest points as the anatomical reference points of the palm.

[0037] The palmprint recognition module is used to analyze the center of gravity position, finger distance pattern, and contour asymmetry of the palm's main outline, and to weight the analysis results to identify whether the palm is left or right.

[0038] It is also used to determine the set of palm print ROIs {ROI_1, ..., ROI_14} based on the main outline of the palm, the position information of the palm anatomical reference points, and the information of the left and right hands, and to perform palm print line detection and filtering within each palm print ROI_i, where ROI_i represents the ROI of the i-th palm print line;

[0039] It is also used to pre-build and train a deep learning semantic segmentation network as a palmprint segmentation model. The trained palmprint segmentation model is used to perform palmprint detection on palm images or palmprint ROI sets to obtain a set of binary masks {Mask_1, ..., Mask_14} that is the same size as the original image and contains 14 channels, where Mask_i represents the segmentation mask of the i-th palmprint line.

[0040] It is also used to perform cross-validation and result fusion on the segmentation mask Mask_i output by the palmprint segmentation model using palmprint ROI_i to obtain a high-confidence palmprint mask Mask_k;

[0041] The multi-dimensional feature quantization module is used to perform multi-dimensional quantization analysis on the high-confidence palmprint mask Mask_k to obtain the palmprint quantization result.

[0042] By adopting the above technical solution, the present invention has the following beneficial effects:

[0043] 1. This invention uniquely provides two technical paths: "anatomical guidance" and "deep learning segmentation". It can make accurate deductions using prior medical knowledge and achieve powerful recognition using deep learning models. The two can be used independently or in combination, which greatly broadens the applicability of the technology.

[0044] 2. This invention achieves true parameter adaptation by dynamically linking the key parameters of the Canny and Hough algorithms to image resolution and ROI size, ensuring detection stability under images of different qualities and sizes. The introduced quality scoring system further improves the accuracy of selecting the best results from multiple candidate targets.

[0045] 3. This invention utilizes palmprint ROIs dynamically generated from anatomical reference points to verify and constrain the output of the palmprint segmentation model, effectively solving errors that may occur in AI models that do not conform to physiological common sense, making the recognition results more reliable and accurate.

[0046] 4. This invention goes beyond simple line detection. Through techniques such as skeletonization and topology analysis, it achieves precise quantification of features that were previously impossible to objectively measure, such as the thickness, curvature, and branching shape of palm lines, providing an unprecedented data dimension for the scientific research of hand diagnosis. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A schematic diagram of an adaptive palmprint multidimensional quantitative analysis method that integrates anatomical features;

[0049] Figure 2 This is a schematic diagram of the anatomical reference point extraction process. (a) is the outline of the palm, (b) is the palm convex hull, and (c) is the interdigital indentation feature point.

[0050] Figure 3 This is a schematic diagram of the palmprint ROI generation results. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] This invention aims to solve the technical problems in existing technologies, such as palm print analysis relying on fixed templates, ignoring anatomical features, having a single quantification dimension, and the lack of medical knowledge constraints in existing AI models leading to poor robustness.

[0053] Specifically, the present invention aims to solve the following technical problems:

[0054] 1. How to provide a dual-path, complementary palmprint recognition framework that can utilize anatomical knowledge for model-driven, accurate searching, and deep learning for powerful data-driven segmentation.

[0055] 2. How to algorithmize abstract anatomical knowledge to generate dynamic, adaptive palmprint regions of interest (ROIs) to accurately guide or verify the palmprint recognition process.

[0056] 3. How to use an innovative fusion mechanism to constrain and correct the output of deep learning models by utilizing anatomical ROIs, thereby significantly improving the accuracy and reliability of recognition.

[0057] 4. How to perform multi-dimensional, high-precision quantification of the identified palm lines, including length, curvature, thickness, and branch topology, to provide comprehensive data support for objective analysis.

[0058] After research, the inventors proposed an adaptive palmprint multi-dimensional quantitative analysis method and system that integrates anatomical features. Its core lies in the deep integration of the anatomical features of the palm with multi-path, adaptive recognition algorithms.

[0059] Embodiment 1 of the present invention provides an adaptive palmprint multi-dimensional quantitative analysis method that integrates anatomical features, comprising the following steps:

[0060] S1. Obtain a palm image, convert the palm image from RGB color space to YCrCb color space, generate a preliminary palm binary mask based on a preset skin color range threshold, and use morphological operations with a large kernel (e.g., 57x57) to remove noise and obtain the final palm mask.

[0061] S2. Extract the outer contour of the final palm mask, select the contour with the largest area as the main contour of the palm, calculate the convex defects between the main contour of the palm and its convex hull, sort the convex defects by depth, and select the deepest points (e.g., 4-8) as the anatomical reference points of the palm.

[0062] S3. Perform center of gravity position analysis, finger distance pattern analysis, and contour asymmetry analysis on the main contour of the palm, and use weighted scoring to identify whether the palm is the left or right hand.

[0063] S4. Determine the set of palm ROIs {ROI_1, ..., ROI_14} based on the main outline of the palm, the position information of the palm anatomical reference points, and the information of the left and right hands. Then, perform palm line detection and screening within each palm ROI_i, where ROI_i represents the ROI of the i-th palm line.

[0064] S41. Based on the main outline of the palm, the position information of the anatomical reference points of the palm, and the information of the left and right hands, determine the horizontal starting boundary, horizontal ending boundary, vertical upper boundary, and vertical lower boundary of each palm line, and generate the ROI of each palm line according to the boundaries of each palm line. Taking the heart line as an example, the generation rule of the heart line ROI is as follows: the outermost point of the palm outline on the little finger side is determined as the horizontal starting boundary of the heart line; the horizontal ending boundary of the heart line is determined according to the palm width ratio and the position of the gap between the index and middle fingers; the position of the palm anatomical reference point with the highest Y coordinate is determined as the vertical upper boundary of the heart line; and the position of the palm midline is determined as the vertical lower boundary of the heart line.

[0065] S42. Preprocess the palm print ROI_i, including: using Gaussian filtering to denoise, using the CLAHE algorithm (e.g., limiting contrast to 2.0) to enhance contrast, and using directional morphological kernels (e.g., horizontal 40x2, vertical 2x40) to enhance palm texture.

[0066] S43. Use the Canny algorithm to perform edge detection on the preprocessed palmprint ROI_i, and use the Hough algorithm to extract line segments. The parameters of the Canny algorithm and the Hough algorithm are dynamically adjusted according to the resolution of the palm image and the size of the palmprint ROI_i.

[0067] S44. Use preset geometric constraints (such as angle range and minimum length ratio) to perform preliminary filtering on all extracted line segments. Use a preset quality scoring system to comprehensively evaluate the three scores of angle deviation, length ratio and position rationality of each candidate line segment, and select the line segment with the highest score as the palmprint recognition result.

[0068] S5. Pre-build and train a deep learning semantic segmentation network as a palmprint segmentation model. Use the trained palmprint segmentation model to perform palmprint detection on the palm image or palmprint ROI set to obtain a set of binary masks {Mask_1, ..., Mask_14} with 14 channels and the same size as the original image. Mask_i represents the segmentation mask of the i-th palmprint line.

[0069] S51. Pre-build a deep learning semantic segmentation network (e.g., based on YOLO11-seg), train it using a large number of palm images annotated by experts, and train the model to perform multi-class segmentation tasks, which can classify each pixel in the image as either background or one of 14 specific palm lines.

[0070] S52. Input the palm image or palm print ROI set into the trained model. The model directly outputs a set of binary masks {Mask_1, ..., Mask_14} that is the same size as the original image and contains 14 channels.

[0071] S6. Cross-validate and fuse the segmentation mask Mask_i output by the palmprint segmentation model using palmprint ROI_i to obtain a high-confidence palmprint mask Mask_k.

[0072] Calculate the intersection-union ratio (IU / R) of the palmprint ROI_i and the corresponding segmentation mask Mask_i, and compare the IU / R with a preset threshold.

[0073] If the crossover ratio is higher than the preset threshold (e.g., 0.5), then the palmprint detection result is confirmed to be a match.

[0074] If the crossover ratio is lower than the preset threshold, it is judged as a false positive and rejected.

[0075] If the traditional method fails to detect the palmprint ROI_i but the palmprint segmentation model does, then the output of the palmprint segmentation model will be adopted as a supplement.

[0076] S7. Perform multi-dimensional quantization analysis on the high-confidence palmprint mask Mask_k to obtain the palmprint quantization results.

[0077] The high-confidence palmprint mask Mask_k is processed using skeletonization algorithms (such as Zhang-Suen thinning) to obtain skeletons S_k with a width of several single pixels;

[0078] The total number of pixels in the skeleton S_k is calculated to obtain the palmprint length. The ratio of the palmprint length to the effective width or height of the palm is calculated to obtain the standardized palmprint length percentage.

[0079] The average and maximum curvature of the palm print are calculated by analyzing the changes in the tangent direction of the skeleton S_k.

[0080] Iterate through each point p on the skeleton S_k, calculate the local tangent direction of the skeleton at point p, and measure the palm print width on the original mask Mask_p along its normal direction.

[0081] The depth of the palm print is quantified by analyzing the grayscale contrast between the palm print area and the adjacent non-palm print area.

[0082] Topological analysis is performed on the skeleton S_k to automatically identify palm print branch points (connection number ≥ 3) and endpoints (connection number = 1). Based on the palm print branch points and endpoints, the total number of palm print branches, the length of each branch, the angle between the trunk and the branch, and the branch density are calculated.

[0083] It should be noted that converting the coordinates of the start and end points and branch points of palm prints to a relative coordinate system based on anatomical benchmarks can achieve a standardized description of their positions.

[0084] Embodiment 2 of the present invention provides an adaptive palmprint multi-dimensional quantitative analysis system that integrates anatomical features to implement the method described in Embodiment 1. The system includes an image acquisition and preprocessing module, an anatomical benchmark extraction module, a palmprint recognition module, and a multi-dimensional feature quantization module.

[0085] The image acquisition and preprocessing module is used to acquire a palm image, convert the palm image from the RGB color space to the YCrCb color space, generate a preliminary palm binary mask based on a preset skin color range threshold, and remove noise using morphological operations with a large kernel to obtain the final palm mask.

[0086] The anatomical reference extraction module is used to extract the outer contour of the final palm mask, select the contour with the largest area as the main contour of the palm, calculate the convex defects between the main contour of the palm and its convex hull, sort the convex defects by depth, and select the deepest points as the anatomical reference points of the palm.

[0087] The palmprint recognition module is used to analyze the center of gravity position, finger distance pattern, and contour asymmetry of the palm's main outline, and to weight the analysis results to identify whether the palm is left or right.

[0088] It is also used to determine the set of palm print ROIs {ROI_1, ..., ROI_14} based on the main outline of the palm, the position information of the palm anatomical reference points, and the information of the left and right hands, and to perform palm print line detection and filtering within each palm print ROI_i, where ROI_i represents the ROI of the i-th palm print line;

[0089] It is also used to pre-build and train a deep learning semantic segmentation network as a palmprint segmentation model. The trained palmprint segmentation model is used to perform palmprint detection on palm images or palmprint ROI sets to obtain a set of binary masks {Mask_1, ..., Mask_14} that is the same size as the original image and contains 14 channels, where Mask_i represents the segmentation mask of the i-th palmprint line.

[0090] It is also used to perform cross-validation and result fusion on the segmentation mask Mask_i output by the palmprint segmentation model using palmprint ROI_i to obtain a high-confidence palmprint mask Mask_k;

[0091] The multi-dimensional feature quantization module is used to perform multi-dimensional quantization analysis on the high-confidence palmprint mask Mask_k to obtain the palmprint quantization result.

[0092] The following specific examples are provided in conjunction with the above embodiments. It should be understood that the following specific examples are only illustrative of the specific implementation of the above embodiments and are not intended to limit the technical solutions of the above embodiments.

[0093] Taking a 2048x1536 pixel, RGB three-channel color image of a left hand uploaded by a user as an example, a multi-dimensional quantitative analysis of palm prints is performed, such as... Figure 1 As shown, the specific process is as follows:

[0094] (I) Image Preprocessing and Anatomical Reference Point Extraction

[0095] 1. Hand image acquisition and preprocessing

[0096] The system receives the user-uploaded image I_color and converts it from the RGB color space to the YCrCb color space. Based on a preset skin tone range threshold (Y[0, 255], Cr[131, 183], Cb[102, 154]), a preliminary handprint binary mask M_skin is generated. A morphological opening operation (using a 57×57 pixel rectangular kernel) is performed on M_skin to remove isolated noise areas outside the wrist, resulting in the final handprint mask M_final.

[0097] 2. Extraction of anatomical reference points

[0098] On M_final, the outermost contour is found using the OpenCV function cv2.findContours. The contour with the largest area is selected as the main contour C of the palm. The convex hull H of contour C is calculated. The convex defects between C and H are calculated using the cv2.convexityDefects function. All convex defects are sorted according to their depth (i.e., the farthest distance from the defect point to the convex hull). The first 4-8 deepest points are selected as the interdigital indentation feature points, i.e., the anatomical reference points of the palm. These points are arranged from left to right and are denoted as points[0] to points[n]. Their pixel coordinates are determined according to the actual detection results. The process is as follows: Figure 2 As shown.

[0099] 3. Hand orientation recognition

[0100] The palm orientation is determined using a multi-feature fusion method.

[0101] Center of gravity position analysis: Calculate the X-coordinate offset of the center of gravity of the palm contour relative to the image center;

[0102] Finger direction analysis: Analyze the pattern of changes in finger distance to identify thumb position features;

[0103] Contour shape analysis: Based on a comprehensive judgment of the asymmetry characteristics of the palm;

[0104] A weighted scoring mechanism is used to determine if the score of the left hand is higher than the set judgment threshold, and the hand is identified as the left hand.

[0105] (II) Palmprint Recognition

[0106] The system starts up in parallel on paths A and B.

[0107] Path A: Anatomy-guided detection

[0108] 1. Adaptive ROI generation

[0109] Taking the left-hand line 1 (heart line) as an example, its ROI is a dynamically generated rectangular area, and the generation rules are as follows:

[0110] The outermost point of the palm contour on the little finger side (search for contour edge points within ±50 pixels of the Y coordinate of the little finger base, and select the outermost edge point as the horizontal starting point) is determined as the horizontal starting boundary (region_left) of the heart line;

[0111] The horizontal termination boundary (region_right) of the heart line is determined based on the proportion of the palm width and the position of the gap between the index and middle fingers (at least 40% of the palm width, but not exceeding the position of the gap between the index and middle fingers, and the horizontal endpoint is selected within this range).

[0112] The location of the highest Y-coordinate anatomical reference point of the palm is determined as the vertical upper boundary (region_top) of the heart line;

[0113] The position of the palm midline (calculated through contour analysis, approximately the midpoint between the base of the fingers and the bottom of the palm) is determined as the vertical lower boundary (region_bottom) of the heart line;

[0114] Generate a rectangular region of interest (ROI) with coordinates ranging from (region_left, region_top) to (region_right, region_bottom), as shown below. Figure 3 As shown.

[0115] 2. Traditional visual inspection

[0116] Within the ROI of the left-hand line 1 (emotional line), a 5×5 Gaussian filter is used to eliminate noise, the contrast is enhanced using the CLAHE algorithm with a contrast limit of 2.0, and directional features are strengthened through morphological operations—the horizontal line is enhanced using a (40, 2) horizontal rectangular kernel, and the vertical line is enhanced using a (2, 40) vertical rectangular kernel.

[0117] The edge detection stage uses the adaptive Canny algorithm: the threshold for high-resolution images (>3000 pixels) is set to [10, 40], the threshold for medium-resolution images (1500-3000 pixels) is set to [20, 80], and the threshold for low-resolution images (<1500 pixels) is set to [30, 120].

[0118] Line segment extraction uses the Hough algorithm with dynamically set parameters: minimum line segment length is 30-50% of the ROI width, maximum connection gap is 3-8% of the ROI width, and the detection threshold is adjusted in the range of 20-60 according to image quality.

[0119] Filter line segments using preset geometric constraints: horizontal line angle ≤ 75°, vertical line angle ≥ 78°, length must be more than 30% of the corresponding dimension of the ROI, and endpoints must be located within the ROI area;

[0120] At the same time, a quality scoring system (angle deviation, length ratio, and position rationality) is introduced to select the best line segment.

[0121] Path B: End-to-end segmentation using deep learning models

[0122] 1. Model execution: Scale the original image I_color to the input size of the palm print segmentation model (e.g., 640x640) and feed it into a trained palm print segmentation model based on YOLO11-seg.

[0123] 2. Output Analysis: The model outputs a prototype mask tensor of [15, 80, 80] and a detection box and mask coefficient tensor of [N, 15+4], where N is the number of detected targets and 15 is the number of categories (14 lines + 1 background). After matrix multiplication and upsampling, a binary mask set {Mask_1, Mask_2, ..., Mask_14} with 14 channels and the same size as the original image (2048x1536) is obtained, where Mask_i is the pixel region of the i-th palmprint line.

[0124] (III) Cross-validation and result fusion

[0125] 1. Verification: Taking line 1 (emotional line) as an example. The system calculates the intersection-over-union ratio (IoU) between Mask_1 obtained from path B and ROI_1 generated from path A.

[0126] Scenario 1 (Matching): The calculated IoU = 0.85. Since 0.85 > 0.5 (preset threshold), the system considers Mask_1 to be a high-confidence correct identification. The final "line 1" result adopts Mask_1 because it contains a more refined curve shape.

[0127] Scenario 2 (Misclassification): Suppose the model incorrectly segments and labels an interfering crease on the wrist as Mask_11 (sexual line). The system calculates the IoU between Mask_11 and ROI_11 and finds that the IoU = 0.05. Since this is far below the preset threshold, the system determines that Mask_11 is a false positive and removes it from the final results.

[0128] 2. Supplement: Suppose that within the ROI of the 4-line (healthy line), the traditional method for path A fails to detect any line segments due to the weak lines, but there is a thin, discontinuous pixel region in Mask_4 of path B. The system adopts Mask_4 as a valid result of the 4-line (healthy line), successfully supplementing the missed detection of the traditional method.

[0129] (iv) Refined Quantization of Multi-Dimensional Features

[0130] The system quantizes each of the high-confidence palmprint masks (such as Mask_1, Mask_4, Mask_13, etc.) obtained after fusion. Taking Mask_13 (Sydney line) as an example:

[0131] 1. Input: A 2048x1536 binary mask Mask_13.

[0132] 2. Skeletonization: Apply the Zhang-Suen thinning algorithm to Mask_13 to obtain a skeleton S_13 with a width of one pixel.

[0133] 3. Length and proportion: Calculate the total number of non-zero pixels in S_13 to obtain the precise length. For example, if the length is 912 pixels and the effective width of the palm is calculated to be 1350 pixels, then the length percentage is (912 / 1350)*100%=67.6%.

[0134] 4. Width: Traverse each point p on the skeleton S_13, calculate the local tangent direction of the skeleton at point p, and measure the width on the original mask Mask_13 along its normal direction. Calculate the width of all points and obtain an average width of 4.1 pixels and a standard deviation of 1.3 pixels (indicating uneven thickness).

[0135] 5. Topology analysis: Use a 3x3 convolution kernel on the skeleton S_13 to count the neighboring pixels. Find a pixel whose neighboring pixels sum to 3 and determine it as a branch point with coordinates (1250, 1080). Calculate the length of the trunk and the branch respectively, and calculate the included angle as 28 degrees.

[0136] (V) Generating Analysis Reports

[0137] The system summarizes the quantitative results of all 14 lines, generates a complete analysis report containing accurate data, and presents it to the user.

[0138] Although the present invention has been disclosed above with reference to embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications and refinements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. An adaptive palmprint multi-dimensional quantitative analysis method integrating anatomical features, characterized in that, Includes the following steps: The process involves acquiring a palm image, converting the palm image from the RGB color space to the YCrCb color space, generating a preliminary palm binary mask based on a preset skin color range threshold, and using morphological operations with a large kernel to remove noise to obtain the final palm mask. Extract the outer contour of the final palm mask, select the contour with the largest area as the main contour of the palm, calculate the convex defects between the main contour of the palm and its convex hull, sort the convex defects by depth, and select the deepest points as the anatomical reference points of the palm. The analysis includes center of gravity position analysis, finger distance pattern analysis, and contour asymmetry analysis of the main contour of the palm. The analysis results are then weighted and scored to identify whether the palm is left or right. Based on the main outline of the palm, the position information of the palm anatomical reference points, and the information of the left and right hands, the set of palm ROIs {ROI_1, ..., ROI_14} is determined, and palm line detection and filtering are performed in each palm ROI_i, where ROI_i represents the ROI of the i-th palm line; A deep learning semantic segmentation network is pre-built and trained as a palmprint segmentation model. The trained palmprint segmentation model is used to detect palmprints in palm images or sets of palmprint ROIs, resulting in a set of binary masks {Mask_1, ..., Mask_14} that is the same size as the original image and contains 14 channels. Here, Mask_i represents the segmentation mask of the i-th palmprint line. The palmprint ROI_i is used to perform cross-validation and result fusion on the segmentation mask Mask_i output by the palmprint segmentation model to obtain a high-confidence palmprint mask Mask_k; Multi-dimensional quantization analysis was performed on the high-confidence palmprint mask Mask_k to obtain the palmprint quantization results.

2. The adaptive palmprint multi-dimensional quantitative analysis method integrating anatomical features according to claim 1, characterized in that, The process involves determining a set of palmprint ROIs {ROI_1, ..., ROI_14} based on the main contour of the palm, the positional information of palm anatomical reference points, and information about the left and right hands. Within each palmprint ROI_i, palmprint line detection and filtering are performed, including: Based on the main outline of the palm, the position information of the palm anatomical reference points, and the information of the left and right hands, determine the horizontal starting boundary, horizontal ending boundary, vertical upper boundary, and vertical lower boundary of each palm line, and determine the ROI of each palm line based on the boundaries of each palm line, thus obtaining the palm line ROI set {ROI_1, ..., ROI_14}. Preprocessing of palm print ROI_i includes: using Gaussian filtering for noise reduction, using the CLAHE algorithm to enhance contrast, and using directional morphology kernels to enhance palm texture; The Canny algorithm is used to perform edge detection on the preprocessed palmprint ROI_i, and the Hough algorithm is used to extract line segments. The parameters of the Canny and Hough algorithms are dynamically adjusted according to the resolution of the palm image and the size of the palmprint ROI_i. All extracted line segments are initially filtered using preset geometric constraints. A preset quality scoring system is used to comprehensively evaluate the angle deviation, length ratio, and position rationality of each candidate line segment. The line segment with the highest score is selected as the palmprint recognition result.

3. The adaptive palmprint multi-dimensional quantitative analysis method integrating anatomical features according to claim 1, characterized in that, The process of cross-validating and fusing the segmentation mask Mask_i output by the palmprint segmentation model using palmprint ROI_i includes: Calculate the intersection-union ratio (IU / R) of the palmprint ROI_i and the corresponding segmentation mask Mask_i, and compare the IU / R with a preset threshold. If the crossover ratio is higher than the preset threshold, the palmprint detection result is confirmed to be a match; If the crossover ratio is lower than the preset threshold, it is judged as a false positive and rejected. If the traditional method fails to detect the palmprint ROI_i but the palmprint segmentation model does, then the output of the palmprint segmentation model will be adopted as a supplement.

4. The adaptive palmprint multi-dimensional quantitative analysis method integrating anatomical features according to claim 1, characterized in that, The multi-dimensional quantitative analysis of the high-confidence palmprint mask Mask_k includes: The high-confidence palmprint mask Mask_k is processed using a skeletonization algorithm to obtain a skeleton S_k with a width of several single pixels. The total number of pixels in the skeleton S_k is calculated to obtain the palmprint length. The ratio of the palmprint length to the effective width or height of the palm is calculated to obtain the standardized palmprint length percentage. The average and maximum curvature of the palm print are calculated by analyzing the changes in the tangent direction of the skeleton S_k. Iterate through each point p on the skeleton S_k, calculate the local tangent direction of the skeleton at point p, and measure the palm print width on the original mask Mask_p along its normal direction. The depth of the palm print is quantified by analyzing the grayscale contrast between the palm print area and the adjacent non-palm print area. Topological analysis is performed on the skeleton S_k to automatically identify palm print branch points and endpoints. Based on the palm print branch points and endpoints, the total number of palm print branches, the length of each branch, the angle between the trunk and the branches, and the branch density are calculated.

5. An adaptive palmprint multi-dimensional quantitative analysis system integrating anatomical features, used to implement the method described in any one of claims 1-4, characterized in that, The system includes an image acquisition and preprocessing module, an anatomical benchmark extraction module, a palmprint recognition module, and a multi-dimensional feature quantization module; The image acquisition and preprocessing module is used to acquire a palm image, convert the palm image from the RGB color space to the YCrCb color space, generate a preliminary palm binary mask based on a preset skin color range threshold, and remove noise using morphological operations with a large kernel to obtain the final palm mask. The anatomical reference extraction module is used to extract the outer contour of the final palm mask, select the contour with the largest area as the main contour of the palm, calculate the convex defects between the main contour of the palm and its convex hull, sort the convex defects by depth, and select the deepest points as the anatomical reference points of the palm. The palmprint recognition module is used to analyze the center of gravity position, finger distance pattern, and contour asymmetry of the palm's main outline, and to weight the analysis results to identify whether the palm is left or right. It is also used to determine the set of palm print ROIs {ROI_1, ..., ROI_14} based on the main outline of the palm, the position information of the palm anatomical reference points, and the information of the left and right hands, and to perform palm print line detection and filtering within each palm print ROI_i, where ROI_i represents the ROI of the i-th palm print line; It is also used to pre-build and train a deep learning semantic segmentation network as a palmprint segmentation model. The trained palmprint segmentation model is used to perform palmprint detection on palm images or palmprint ROI sets to obtain a set of binary masks {Mask_1, ..., Mask_14} that is the same size as the original image and contains 14 channels, where Mask_i represents the segmentation mask of the i-th palmprint line. It is also used to perform cross-validation and result fusion on the segmentation mask Mask_i output by the palmprint segmentation model using palmprint ROI_i to obtain a high-confidence palmprint mask Mask_k; The multi-dimensional feature quantization module is used to perform multi-dimensional quantization analysis on the high-confidence palmprint mask Mask_k to obtain the palmprint quantization result.

Images