An AI-based facial skin texture feature extraction method and system

Abstract

This invention relates to the field of feature extraction technology, specifically to an AI-based method and system for facial skin texture feature extraction. The method includes the following steps: using an image acquisition device, a face region is captured and grayscale spatial distribution data is obtained; grayscale differences and transition features of each pixel in multiple directions are analyzed; and the responses of each directional channel are fused to obtain the skin texture distribution result. In this invention, the input stage maintains a balanced foundation for the original data through brightness spatial consistency and pixel continuity constraints. During feature extraction, multi-directional grayscale change aggregation and transition trajectory analysis are employed to distinguish the texture direction and structural distribution characteristics of local areas. For areas with significant differences, the allocation of parameters for each channel is adjusted based on the dominant directional response. Reflection interference positions are handled through dynamic weight correction, ensuring that the output result possesses structural integrity and directional response coherence in real space, effectively improving the skin texture mapping capability in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of feature extraction technology, and in particular to an AI-based method and system for extracting facial skin texture features. Background Technology

[0002] Feature extraction involves extracting representative feature information from raw data and is widely used in image recognition, speech recognition, signal processing, biometrics, and other fields. It includes feature selection, feature transformation, feature dimensionality reduction, and feature representation. Traditional facial skin texture feature extraction methods typically involve acquiring facial images using image acquisition devices and then employing image processing algorithms such as gray-level co-occurrence matrix, wavelet transform, and local binary pattern to extract skin surface texture information, which is then used for feature modeling.

[0003] Existing technologies have limited adaptability to lighting conditions when collecting raw data. Under conventional processes, pixel brightness changes are significantly affected by ambient light distribution, making it difficult to uniformly calibrate. Feature construction based on fixed operators cannot take into account both directional details and regional continuity, resulting in a one-sided representation of structural information in the skin region. Texture direction exhibits breaks or false splicing in spatial representation. Local reflections or noise factors can also mask real texture changes, further impairing the directional expressive ability of feature extraction results. The response to microstructures is unstable, and overall, it cannot provide a continuous and reliable feature foundation for subsequent analysis stages. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides an AI-based method and system for extracting facial skin texture features. The technical solution is as follows: On the one hand, an AI-based method for extracting facial skin texture features is provided, including the following steps: S1: Based on the image acquisition device, the ambient brightness distribution is recorded through the light source sensor, the shooting angle and exposure parameters are adjusted, effective pixels are screened according to the standard, the uniformity of grayscale pixel brightness is evaluated, the continuous area of ​​pixel distribution is determined, and grayscale spatial distribution data is obtained. S2: Based on the grayscale spatial distribution data, compare the grayscale differences of pixels in multiple directions, analyze the changing trend using a sliding window, determine the jump coordinates by combining the skin high contrast feature law, identify the feature display channel, and obtain a multi-directional gradient response sequence. S3: Based on the multi-directional gradient response sequence, determine the dominant direction change trend, compare the gray level difference between the dominant direction and each direction, and identify pixels with inconsistent trends by combining the neighborhood difference distribution to obtain the dominant direction change identifier set; S4: Analyze the feature distribution data in the set of dominant directional change identifiers, identify the reflection enhancement edge region, compare the grayscale curves of the dominant and adjacent channels based on the adjacent channel data, identify interfering pixels and adjust the weight allocation to obtain the directional channel weight configuration. S5: Based on the directional channel weight configuration, map and fuse the response to the original coordinates, screen for areas with strong reflection in the ambient light data, evaluate the structural correspondence, and integrate the channel response values ​​to obtain the skin texture distribution results.

[0005] On the other hand, the grayscale spatial distribution data includes a brightness distribution map, pixel continuity and regional uniformity parameters, the multi-directional gradient response sequence includes a direction change data set, jump region information and channel response feature set, the direction dominant change identifier set includes a dominant direction label, trend differentiation mark and feature pixel identifier, the direction channel weight configuration includes a main channel weight parameter, weight allocation index and interference region weight data, and the skin texture distribution result includes a direction distribution map, texture partition information and intensity expression matrix.

[0006] On the other hand, the specific steps for obtaining the grayscale spatial distribution data are as follows: S101: Based on the image acquisition device, analyze the ambient brightness distribution information collected by the light source detection sensor, determine the trend of light change in each area, calculate the brightness continuity between adjacent pixels, and statistically analyze the balance state between pixel areas according to the spatial arrangement change law to obtain spatial brightness distribution data. S102: Based on the spatial brightness distribution data, compare the spatial distribution of pixel grayscale, filter the grayscale data of all pixels in the image frame, judge the validity of pixel information according to the exposure reference standard, filter out pixels that do not meet the reference standard, and obtain a set of qualified grayscale pixels. S103: Based on the set of qualified grayscale pixels, analyze the grayscale distribution uniformity of the pixels, use a sliding window to gradually compare the coherence of grayscale changes in the region, determine the continuous aggregation trend of pixels in the space, optimize the spatial structure of pixel distribution, and obtain grayscale spatial distribution data.

[0007] On the other hand, the steps for obtaining the multi-directional gradient response sequence are as follows: S201: Based on the grayscale spatial distribution data, determine the grayscale level changes of pixels in the horizontal, vertical and diagonal directions, compare the grayscale level changes of adjacent pixels in turn through a sliding window, calculate the continuous change features of pixel sequences in different directions, filter out directional distributions with cumulative change trends, and obtain multi-directional gradient feature groups. S202: Based on the multi-directional gradient feature group, compare the magnitude of gray level changes in each direction, determine the direction in which the pixel's gray level changes are prominent in multiple directions, locate the boundary region according to the pixel's change sequence in each direction, filter the pixel coordinates of gray level changes, and obtain the jump pixel index set. S203: Based on the jump pixel index set, analyze the directional change order of the corresponding pixels, compare the gray level change characteristics of each direction, determine the spatial distribution of gray level changes between channels, integrate the channel characteristics of gray level change regions in different directions, and obtain a multi-directional gradient response sequence.

[0008] On the other hand, the steps for obtaining the direction-dominant change identifier set are as follows: S301: Based on the multi-directional gradient response sequence, determine the gray level change trend of pixels in each directional channel, statistically analyze the gray level change trend of pixels in each directional channel, compare the continuity and offset direction of gray level change in each channel, and obtain the dominant channel sorting information. S302: Based on the dominant channel sorting information, compare the grayscale change amplitude of the dominant direction with that of other directional channels, calculate the cumulative offset of grayscale level differences between channels, determine the coordination of changes between channels in each direction, identify the pixel region where the grayscale distribution difference of the channels is concentrated, and obtain the directional difference feature index. S303: Based on the directional difference feature index, determine the relationship between the dominant direction and the directional distribution of adjacent pixels, analyze the trend of change of the dominant direction and the distribution of differences in surrounding pixels, screen pixels whose directional change trend differs from that of the neighborhood, integrate pixel and dominant direction information, and obtain a set of dominant directional change identifiers.

[0009] On the other hand, the specific steps for obtaining the directional channel weight configuration are as follows: S401: Based on the dominant direction change identifier set, determine the spatial distribution and directional feature continuity of pixels within the marked area, filter edge pixels that exhibit illumination enhancement in adjacent channels, statistically analyze the spatial combination relationship of pixels, and obtain reflection edge channel distribution data. S402: Based on the reflection edge channel distribution data, compare the grayscale distribution curves of the dominant direction channel and adjacent channels, analyze the differences in grayscale distribution trends between adjacent channels, screen pixel combinations in which grayscale trends change, and obtain an interference pixel identification list. S403: Determine the grayscale response distribution of the channels involved in the interference pixel identification list, analyze the proportional relationship of weight changes between the dominant channel and adjacent channels, adjust the weight allocation parameters of the dominant direction and adjacent direction channels, and obtain the directional channel weight configuration.

[0010] On the other hand, the specific steps for obtaining the skin texture distribution results are as follows: S501: Based on the directional channel weight configuration, analyze the adjusted channel response data, determine the response characteristics of pixel arrangement in each channel, optimize the correspondence between channel direction and original pixel coordinates, calculate the response trend of each channel at the same coordinate point, and obtain the channel fusion response sequence. S502: Based on the channel fusion response sequence, compare the spatial distribution differences with the ambient light reference data, analyze the correlation between the fusion grayscale response of each pixel and the ambient light change, filter the pixels with grayscale trend changes in multiple channels, statistically analyze the spatial aggregation features, and obtain the pixel index of the reflection area. S503: Based on the pixel index of the reflection area, analyze the directional correspondence between the multi-channel fusion response and the skin partition structure, screen the areas that meet the partition texture direction, aggregate the partition direction data, and obtain the skin texture distribution result.

[0011] On the other hand, the ambient brightness distribution data refers to the light intensity distribution at each location in the entire acquisition scene obtained by the light source detection sensor, and the effective pixels refer to the set of pixels that meet the conditions after screening for exposure, occlusion and noise conditions in the acquired original image.

[0012] On the other hand, the high-contrast skin features refer to the areas in facial skin images where local grayscale changes and bright-dark contrasts are prominent due to pores, wrinkles, and acne scars. The enhanced reflection edge areas refer to the local edge areas in facial images where brightening and specular reflection phenomena are caused by skin oil, moisture, or light reflection.

[0013] On the other hand, an AI-based facial skin texture feature extraction system is provided, which is applied to an AI-based facial skin texture feature extraction method, including: The illumination sensing module is based on an image acquisition device. It records the ambient brightness distribution through a light source sensor, adjusts the shooting angle and exposure parameters, screens effective pixels according to standards, evaluates the uniformity of grayscale pixel brightness, determines the continuous area of ​​pixel distribution, and obtains grayscale spatial distribution data. The gradient extraction module compares the grayscale differences of pixels in multiple directions based on the grayscale spatial distribution data, analyzes the changing trend using a sliding window, judges the jump coordinates by combining the high contrast feature of skin, identifies the feature display channel, and obtains a multi-directional gradient response sequence. Based on the multi-directional gradient response sequence, the direction recognition module determines the trend of dominant direction change, compares the grayscale difference between the dominant direction and each direction, and identifies pixels with inconsistent trends by combining the neighborhood difference distribution to obtain a set of dominant direction change identifiers. The weight adjustment module analyzes the feature distribution data of the dominant change identifier set, identifies the reflection enhancement edge region, compares the grayscale curves of the dominant and adjacent channels based on the adjacent channel data, identifies interfering pixels and adjusts the weight allocation to obtain the directional channel weight configuration. The texture fusion module maps and fuses responses to the original coordinates based on the directional channel weight configuration, screens areas with strong reflection for ambient light data, evaluates structural correspondence, and integrates channel response values ​​to obtain skin texture distribution results.

[0014] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: During the input phase, the original data is kept balanced by constraining the spatial consistency of brightness and the continuity of pixels. In the feature extraction process, multi-directional gray-level change aggregation and jump trajectory analysis are used to distinguish the texture direction and structural distribution characteristics of local regions. For areas with significant differences, the parameter allocation of each channel is adjusted in combination with the dominant directional response. The position of reflection interference is handled by dynamic weight correction. The fusion mapping of features in each direction is realized at the pixel level, so that the output results have structural expression integrity and directional response coherence in real space, effectively improving the skin texture mapping capability in complex scenes. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0016] Figure 1 This is a flowchart of the main steps of the present invention; Figure 2 This is a flowchart of steps S1 of the present invention; Figure 3 This is a flowchart of steps S2 of the present invention; Figure 4 This is a flowchart of steps S3 of the present invention; Figure 5 This is a flowchart of step S4 of the present invention; Figure 6 This is a flowchart of steps S5 of the present invention; Figure 7 This is a system block diagram of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0019] This invention provides an AI-based method for extracting facial skin texture features, such as... Figure 1 As shown, it includes the following steps: S1: Based on the image acquisition device, the ambient brightness distribution data is recorded using the light source detection sensor. The shooting angle and exposure parameters are adjusted for the face area. Valid pixels are screened according to the exposure reference standard. The acquired grayscale pixels are evaluated according to the uniformity of brightness distribution to determine the continuous area of ​​pixel distribution and obtain grayscale space distribution data. S2: Based on grayscale spatial distribution data, compare the grayscale differences of each pixel in the horizontal, vertical and diagonal directions, use a sliding window to analyze the continuous grayscale change trend in the four directions point by point, combine the directional distribution law of the high contrast feature area of ​​the skin, determine the coordinates of the jump area, and identify the channel data that shows the directional change feature to obtain the multi-directional gradient response sequence. S3: Based on the multi-directional gradient response sequence, determine the gray-scale change trend of the dominant directional channel of each pixel. Using the difference magnitude sorting method, compare the gray-scale difference between the dominant direction and other directions one by one. Simultaneously combine the difference distribution data of each direction in the pixel neighborhood to identify pixels whose dominant directional change trend is inconsistent with the surrounding pixels, and obtain the dominant directional change identifier set. S4: Analyze the feature distribution data of the dominant change identifier set, identify the pixel channel structure corresponding to the reflection enhancement edge area, compare the gray distribution curves of the dominant channel and its adjacent channels based on the gray change data of adjacent channels, identify the pixels with interference effects in the directional channel, and adjust the weight allocation value of each channel to obtain the directional channel weight configuration. S5: Based on the channel response data after the directional channel weight configuration is adjusted, the response sequence after channel fusion is mapped to the original pixel distribution coordinates. Strong reflection areas are screened for ambient light reference data. The structural correspondence between the fused responses of each direction and the skin partition is evaluated. The directional channel response values ​​are integrated point by point at the pixel level to obtain the skin texture distribution results.

[0020] The grayscale spatial distribution data includes brightness distribution map, pixel continuity and regional uniformity parameters; the multi-directional gradient response sequence includes directional change data set, jump region information and channel response feature set; the dominant directional change identifier set includes dominant directional label, trend differentiation mark and feature pixel identifier; the directional channel weight configuration includes main channel weight parameter, weight allocation index and interference area weight data; and the skin texture distribution results include directional distribution map, texture partition information and intensity expression matrix.

[0021] In S1, ambient brightness distribution data refers to the distribution of light intensity at various locations within the entire acquisition scene, obtained through a light source detection sensor or imaging system. This reflects the spatial uniformity and brightness variations of the ambient light, and is used for adaptive adjustment of subsequent shooting parameters. Exposure reference standards refer to technical specifications or standards used to determine whether pixels are within a reasonable exposure range. These can be set based on the device's built-in exposure baseline, industry standards, or experience gained from acquisition in specific scenarios, ensuring that the image signal is not overexposed or underexposed. Effective pixels refer to those pixels in the acquired raw image that, after screening for exposure, occlusion, noise, and other conditions, meet quality requirements and can be used for subsequent analysis. The set of pixels processed; grayscale pixels refer to single-channel grayscale pixels obtained after processing the original RGB image through a grayscale transformation algorithm (such as weighted averaging), mainly used for subsequent image structure and texture analysis; uniformity of brightness distribution refers to the degree of consistency of the brightness distribution of grayscale image pixels in space, reflecting the overall brightness balance of the image, and is a reference for judging image quality and the usability of subsequent data; continuous pixel distribution region refers to a set of pixels with similar brightness, adjacent to each other and without abrupt changes in grayscale space, usually represented as a facial skin segment with complete structure and continuous boundaries, which is the basic data block for subsequent feature extraction.

[0022] In S2, continuous grayscale change trend refers to the gradual increase or decrease of pixel grayscale along a certain direction (horizontal, vertical, diagonal), reflecting the extensibility of skin texture and local light and shadow structure; point-by-point analysis refers to performing data analysis operations on each pixel and its neighborhood one by one, without overall averaging, but tracking the grayscale changes and feature states of each point in detail; high-contrast skin feature area refers to the area in the facial skin image where local grayscale changes are drastic and the contrast between light and dark is prominent due to pores, wrinkles, acne marks, etc., and is the main area carrying the differences in skin structure and texture features; directional distribution law refers to the spatial distribution trend and characteristics of texture or light and dark changes along different directions (such as skin fiber arrangement, hair growth direction) within a local area; jump region coordinates refer to the position of obvious grayscale change in the directional grayscale sequence, which, after being located by coordinates, helps to judge key parts such as the edge and boundary of skin texture; channel data refers to the set of numerical values ​​related to texture changes obtained by analysis in different directions (or feature channels), with each direction being a data channel, used for multi-dimensional joint description of skin features.

[0023] In S3, the difference magnitude sorting method refers to sorting the grayscale difference of the same pixel in multiple directions to identify the dominant (largest change) and secondary (smallest change) directions, which serve as the main feature extraction method. The difference distribution data in each direction refers to the set of grayscale changes of each pixel in multiple directions such as horizontal, vertical, and diagonal, reflecting the anisotropic characteristics of the texture. The dominant direction change trend refers to the spatial continuity and change pattern of the direction with the most obvious grayscale change (largest ranking), which is an important reference for determining the main direction of skin fine structure or texture.

[0024] In S4, the enhanced reflection edge region refers to the local edge area in the facial image that is bright and specularly reflective due to skin oil, moisture, or light reflection, which is prone to causing texture feature breakage or loss; adjacent channels refer to the channels that are close to each other in space or direction when analyzing channel data in a certain dominant direction (such as adjacent to the upper, lower, left, right, and diagonal channels in the horizontal direction); grayscale distribution curve refers to the numerical curve formed by the change of pixel grayscale with position along a certain direction, used to compare the consistency and change pattern of texture distribution in different directions or regions; interference effect refers to abnormal texture signals caused by non-skin body feature factors such as reflection, occlusion, and image noise, which need to be detected and corrected by algorithms.

[0025] In S5, the response sequence after channel fusion refers to the response data sequence of each pixel in the multi-channel feature space after multi-directional channel weight adjustment and channel data fusion, which is used to construct the final texture representation; ambient light reference data refers to the auxiliary lighting information related to the initial illumination distribution, used for subsequent analysis and reflection area screening, which is usually provided by the sensor or algorithm module during acquisition; strong reflection area refers to the set of pixels that are locally bright and have obvious brightness changes with the surrounding area under the influence of ambient light, usually corresponding to areas such as skin reflection and oiliness; structural correspondence refers to the degree of fit between the channel fusion response sequence and the spatial and texture structure of the actual skin partition (such as T-zone, U-zone, nasal wing, etc.), which is used to evaluate the rationality of the model output.

[0026] like Figure 2 As shown, the specific steps for obtaining grayscale spatial distribution data are as follows: S101: Based on the image acquisition device, analyze the ambient brightness distribution information collected by the light source detection sensor, determine the trend of light change in each area, calculate the brightness continuity between adjacent pixels, and statistically analyze the balance state between pixel areas according to the spatial arrangement change law to obtain spatial brightness distribution data. The illumination intensity values ​​between adjacent pixels are extracted in the horizontal and vertical directions of the image. For each group of adjacent pixels, the illumination difference between them is calculated sequentially. If the difference between three adjacent pixels is less than 15, the pixel column or row is considered a brightness continuous region. This process is repeated across the entire image frame to construct a brightness change trend map for each region. Based on this, the illumination values ​​of all pixels in each local region (e.g., a 50×50 pixel block) are further statistically analyzed, and the overall fluctuation amplitude within the region is calculated. If the fluctuation amplitude is less than 10, the region is marked as a brightness stable region. Subsequently, the regions are summarized according to their proportion to the total pixel area of ​​the image. When the proportion of stable regions exceeds 40% of the total image area, the brightness is considered stable. The system records the center point coordinates and average brightness value of the region and uses it as the illumination equalization region for subsequent processing. For example, in a 640×480 image, if 20 50×50 regions are detected that meet the brightness stability standard, the cumulative pixel area is 50,000 pixels, accounting for about 16% of the total image area. At this time, it does not meet the set 40% stable region ratio, so it will not be judged as an illumination equalization image. Conversely, if the cumulative stable region reaches 50, it is 125,000 pixels, accounting for about 40.7%. Only then does it meet the set condition and can be marked as an illumination equalization image. By statistically analyzing the illumination continuity, brightness stability, and spatial distribution of each region in the entire image, spatial brightness distribution data is obtained.

[0027] S102: Based on spatial brightness distribution data, compare the spatial distribution of pixel grayscale, filter the grayscale data of all pixels in the image frame, judge the validity of pixel information according to the exposure reference standard, filter out pixels that do not meet the reference standard, and obtain a set of pixels with qualified grayscale. The grayscale value of each pixel in the image is extracted one by one, and the difference between this difference and the average brightness of its surrounding area is calculated. If the difference is less than or equal to 20, the pixel is considered to have no significant brightness shift. Simultaneously, the pixel's own grayscale value is assessed; it should fall within the set exposure reference range, i.e., between 60 and 200. If it is below 60, it is considered underexposed; if it is above 200, it is considered overexposed. Such pixels are directly discarded. Next, the grayscale values ​​of the eight neighboring pixels in the image's region are extracted, and the average grayscale value of the neighborhood is calculated. If the average neighborhood value also falls within the range of 60 to 200... Furthermore, if the difference between the grayscale value of a pixel and the center pixel does not exceed 25, the pixel is further confirmed as valid. For example, if a pixel has a grayscale value of 145, the average brightness of its area is 135, the difference is 10, which is less than the set value of 20, the average grayscale value of its neighboring area is 152, the difference is 7, which is less than 25, and all grayscale values ​​fall within the valid range, then the pixel is retained as a qualified grayscale pixel. Otherwise, it is removed from the subsequent analysis. This processing logic covers the entire image, completing the screening and judgment point by point, forming a set of qualified grayscale pixels composed of multiple pixels with grayscale values ​​within a reasonable exposure range and a balanced lighting area.

[0028] S103: Based on the set of qualified grayscale pixels, analyze the grayscale distribution uniformity of pixels, use a sliding window to gradually compare the coherence of grayscale changes in the region, judge the continuous aggregation trend of pixels in the space, optimize the spatial structure of pixel distribution, and obtain grayscale spatial distribution data. A sliding window with a size of 5×5 pixels is set on the entire image, and it slides sequentially to cover the image area with a step size of 1 pixel. In each window, the grayscale values ​​of 25 pixels are extracted and statistically analyzed. First, the overall fluctuation range is calculated. When the value is less than 12, the grayscale distribution within the window area is considered relatively uniform. Then, the directional continuity is judged. Taking the center pixel of the window as the reference, adjacent pixels in the left-right, top-bottom, and two diagonal directions are taken, and the grayscale difference change in each direction is calculated. If the grayscale difference sequence changes continuously in a certain direction by less than 30, that is, the sum of the absolute values ​​of the two differences is considered a valid value. If the number of pixels is less than 30, the direction is considered a continuous direction of grayscale change. At the same time, the number of qualified grayscale pixels in the window must be no less than 20, accounting for more than 80% of the total number of pixels in the window. If the above conditions are met, the position of the center pixel of the window is recorded as a valid position in the continuous grayscale structure region. For example, in a certain window, there are 22 valid pixels, the overall grayscale fluctuation is 8, the horizontal grayscale difference is -6 and 4, and the sum of the absolute values ​​is 10. The continuity condition is met, so the region is marked as a stable grayscale space region. All the center pixels of the windows that meet the conditions are set into a new structural distribution region, and the grayscale space distribution data is output.

[0029] like Figure 3 As shown, the specific steps for obtaining the multi-directional gradient response sequence are as follows: S201: Based on grayscale spatial distribution data, determine the grayscale level changes of pixels in the horizontal, vertical and diagonal directions, compare the grayscale level changes of adjacent pixels in turn through a sliding window, calculate the continuous change features of pixel sequences in different directions, filter out directional distributions with cumulative change trends, and obtain multi-directional gradient feature groups. A fixed-size sliding window (5×5) is established centered on each pixel in the image. Gray-level sequences of five adjacent pixels are extracted through this window in the horizontal, vertical, and two diagonal directions. The gray-level difference between adjacent pixels is calculated sequentially. If the absolute value of the gray-level difference between adjacent pixels is greater than or equal to 15 at three consecutive points, a continuous gray-level change trend is identified in that direction. This judgment is performed in each direction, and the number of gray-level change trends at each pixel position is counted. Based on the image distribution pattern, a threshold of 3 for the number of consecutive directional changes is set. That is, if three consecutive gray-level changes in any direction exceed the set value, that direction is determined to have cumulative change characteristics. For example, in a certain image... The horizontal grayscale sequence of a pixel is [120, 137, 152, 168, 179], with grayscale differences of 17, 15, 16, and 11 respectively. The first three differences are all greater than 15, which meets the trend condition. The vertical grayscale sequence is [130, 132, 129, 128, 127], with small fluctuations in grayscale differences, and the difference is within 5, which does not meet the condition. Therefore, this pixel is only recorded as a change direction in the horizontal direction. By traversing all qualified pixels in the image, a structure consisting of direction labels and corresponding positions is formed. The change state of each pixel in four directions is recorded, and the combination of pixels that meet the trend judgment criteria and their directions is statistically analyzed, thereby obtaining the multi-directional gradient feature group in the entire image.

[0030] S202: Based on multi-directional gradient feature groups, compare the magnitude of gray level changes in each direction, determine the direction in which the pixel's gray level changes are prominent in multiple directions, locate the boundary region according to the pixel's change sequence in each direction, filter the pixel coordinates of gray level changes, and obtain the jump pixel index set. For each pixel, extract the set of continuous grayscale changes in four directions. Calculate the difference between the maximum and second-largest changes in each direction. If the difference exceeds a set threshold of 20, the direction of maximum change is identified as the primary change direction for that pixel and marked. Then, perform gradient accumulation on the continuous grayscale differences along the primary change direction. If the total change among five pixels exceeds 50, the pixel is marked as a boundary candidate pixel, and its coordinates in the image are recorded. For example, if a pixel's cumulative grayscale changes in the horizontal, vertical, left diagonal, and right diagonal directions are 45 and 18 respectively... The values ​​are 12, 9, with a maximum value of 45 and a second maximum value of 18. The difference is 27, which is greater than the set threshold of 20. Therefore, the horizontal direction is the main change direction, and its cumulative change is 45, which does not meet the boundary judgment requirements. If another pixel has a cumulative change of 60 in the horizontal direction, which meets the boundary conditions, its coordinates [120, 185] are written into the jump index list. The system traverses each pixel in the image that meets the feature group conditions, completes the selection of the main change direction, the judgment of the cumulative change value, and the screening of the jump area, and forms a jump pixel index set. This index set contains the position coordinates of each mutated pixel and the direction label corresponding to the mutation.

[0031] S203: Based on the jump pixel index set, analyze the directional change order of the corresponding pixels, compare the gray level change characteristics of each direction, determine the spatial distribution of gray level changes between channels, integrate the channel features of gray level change regions in different directions, and obtain a multi-directional gradient response sequence. Extract the grayscale change sequence of each pixel in four directions, construct a grayscale difference sequence for each direction, and compare the change magnitude and direction of the sequence to see if they exhibit increasing, decreasing, or reversing characteristics. If the change sequence in a certain direction shows a sudden change in direction in three consecutive sets of grayscale difference values, for example, the grayscale difference changes from +20 to -15 and then to +17, it is determined that there is an inconsistent change in that direction. Conversely, if the difference values ​​continuously show the same trend, such as +12, +16, +19, it is a consistent change, and the former is marked as an abnormal change direction. At the same time, the change pattern of each pixel in different directions is statistically analyzed within the jump direction. The system groups and categorizes channel data that exhibit similar change patterns in all directions. For example, if pixel A has similar grayscale change trends in the horizontal and left diagonal directions (both continuously increasing) and a decreasing trend in the vertical direction, then the horizontal and left diagonal directions are integrated into the same directional channel. A grayscale change record list is established for each abruptly changing pixel based on the channel direction. After the channel change structure of all abruptly changing pixels is completed, the change sequences of channels in the same direction are merged to construct a grayscale change mapping structure under multiple pixels and multiple directions, forming a set of grayscale response sequences of all abruptly changing pixels in each direction, called the multi-directional gradient response sequence.

[0032] like Figure 4As shown, the specific steps for obtaining the dominant direction change identifier set are as follows: S301: Based on the multi-directional gradient response sequence, determine the gray level change trend of pixels in each directional channel, statistically analyze the gray level change trend of pixels in each directional channel, compare the continuity and offset direction of gray level change in each channel, and obtain the dominant channel sorting information. The grayscale change sequence of a pixel is read pixel by pixel in the four channels of horizontal, vertical, left diagonal, and right diagonal directions. The grayscale difference between adjacent pixels in each direction is counted, and the increase and decrease trend of grayscale values ​​is recorded sequentially. By calculating the number of consecutive positive and negative changes in each direction and comparing the length of their continuous segments, if the number of consecutive increases and decreases in a certain direction exceeds a set standard threshold of 3 times, and the length of the maximum continuous change segment is not less than 5 pixels, then that direction is marked as having a grayscale change trend. Then, the grayscale trends of the marked pixels in the four directions are compared horizontally, and the total number of grayscale change trends and the directional offset of positive and negative changes in each channel are counted. For example, if the grayscale sequence of a pixel in the horizontal channel is [122, 130, 137, 145, 153], the consecutive increases are 8, 7, 8, 8, and the total length of the continuous increase segment is 4. If the judgment condition is met, it is marked as a stable upward trend. The vertical sequence is [150, 142, 135, 129, 122]. The direction of change is downward. The length of the continuous segment is also 4. The two directions of change are opposite. The diagonal change is irregular, with only two continuous changes, which do not constitute a trend direction. The horizontal and vertical directions are marked as the main trend directions. Then, the number of trend direction types appearing in each channel for each pixel is counted, and the grayscale fluctuation values ​​of each channel are sorted. The grayscale fluctuation value is the sum of the absolute values ​​of all grayscale differences in the channel. The channel with the largest grayscale fluctuation is sorted first. For example, if the total fluctuation in the horizontal direction is 32, the vertical direction is 28, and the other directions are less than 20, then the channel sorting result is horizontal first, vertical second, left diagonal third, and right diagonal fourth. The channel sorting information of each pixel is recorded in a matrix structure, and the dominant channel sorting information is output.

[0033] S302: Based on the dominant channel sorting information, compare the grayscale change amplitude of the dominant direction with that of other directional channels, calculate the cumulative offset of grayscale level differences between channels, determine the coordination of changes between channels in each direction, identify the pixel region where the grayscale distribution difference of the channels is concentrated, and obtain the directional difference feature index. Extract the grayscale difference sequence of each pixel in the dominant direction and other directional channels. Calculate the grayscale difference at each corresponding position between the dominant direction channel and other directional channels, and accumulate these differences. If the total accumulated grayscale difference exceeds a set offset threshold of 50, the dominant channel change of the pixel is determined to be significantly deviated from other directions. The stability of the offset is then judged based on the continuity of the grayscale difference change. If there are at least three consecutive positions with grayscale differences greater than 15, it is considered a region with poor change coordination. Record the coordinates of the pixel and its dominant channel direction label in the image. Simultaneously, count the number of consecutive pixel regions exceeding a certain area in the entire image (set to be 100 pixels or more). If the channel grayscale difference of multiple consecutive pixels exceeds the set range, and the direction label... If they are consistent, they are classified as concentrated areas of grayscale distribution difference. For example, if a pixel's dominant direction is vertical, and its vertical grayscale difference is 10, 12, 14, 16, 18, its directional difference is 4, 5, 7, 6, 5, and the difference at each position is 6, 7, 7, 10, 13, with a total difference of 43, which is less than the set threshold. The consistency is considered acceptable, and it is not recorded as a difference area. Another pixel's dominant direction is horizontal, and its horizontal grayscale change is 14, 17, 19, 21, 23, with only directional differences of 2, 3, 5, 4, 5. The total difference is 80, and the difference at three consecutive points exceeds 15. This pixel is then recorded as a point of concentrated difference. By scanning the entire image, the set of pixel coordinates with significant directional difference features is selected, and the directional difference feature index is output.

[0034] S303: Based on the orientation difference feature index, determine the relationship between the dominant orientation and the orientation distribution of adjacent pixels, analyze the trend of the dominant orientation change and the difference distribution of surrounding pixels, screen pixels whose orientation change trend differs from the neighborhood, integrate pixel and dominant orientation information, and obtain the orientation dominant change identifier set. Extract the dominant direction label of each labeled pixel and construct a 3×3 neighborhood window around it. Read the dominant direction information and grayscale change trend of other pixels in the neighborhood sequentially. If the dominant direction of a pixel is inconsistent with the direction labels of more than five pixels in the neighborhood, and its grayscale change trend differs from the average trend of the neighborhood by more than 20, then the pixel is marked as a directional trend anomaly. Further, a unified judgment is made on the grayscale changes of all pixels within the neighborhood of this pixel. If at least four directional change trends are similar within the neighborhood, and the trend direction of this pixel differs from these four pixels by more than 20, then the pixel is determined as a directional inconsistency point. For example, if the current pixel's dominant direction is vertical, the change trend is continuously decreasing, and the total change amplitude is 50, and five pixels in the neighborhood have a dominant direction of horizontal, a change trend is continuously increasing, and the average change amplitude is 45, then the number of inconsistent directions is greater than five, and the trend change directions are opposite, showing a significant difference. In this case, the pixel coordinates are recorded in the directional dominant change identifier set. Simultaneously, its dominant direction, trend curve type, change amplitude, and other information are integrated to form a complete identifier item. All such identifier items constitute the directional dominant change identifier set.

[0035] like Figure 5 As shown, the specific steps for obtaining the directional channel weight configuration are as follows: S401: Based on the orientation-dominant change identifier set, determine the spatial distribution and orientation feature continuity of pixels within the marked area, filter edge pixels that exhibit illumination enhancement in adjacent channels, and statistically analyze the spatial combination relationship of pixels to obtain reflection edge channel distribution data; The image coordinates of the marked pixels are located and their dominant directions are calculated. For each marked point, a 3×3 neighborhood window is constructed to read the information of its surrounding pixels, recording the dominant direction and grayscale value of each pixel in the neighborhood. Simultaneously, the continuity of the dominant direction of the pixel is calculated. If the direction of the central pixel is consistent with the direction of more than 5 surrounding pixels, and the grayscale change direction is the same, then the region is determined to have directional continuity. Further, the directional channel information adjacent to the dominant direction of the central pixel is extracted within this neighborhood, and its grayscale value is compared. If there are ≥3 pixels in the adjacent channels with a grayscale value greater than 20 of the central pixel, then the neighborhood is marked in memory. The illumination enhancement edge features are analyzed, and the direction of the enhanced channel is recorded. For example, when the gray value of the center pixel is 135, the gray values ​​of the pixels in its adjacent channels are 158, 162, 161, 140, and 133. Among them, three values ​​are greater than the center value by more than 20, which meets the illumination enhancement judgment condition. The edge where the pixel is located is marked as a reflection edge. Then, the center position of the pixel, the dominant direction, and the corresponding enhancement channel direction are stored together as a structure, and its spatial position and directional channel mapping relationship are recorded. The above operation is repeated on the entire image to extract all edge pixels that meet the enhancement condition and their adjacent directional relationships, forming reflection edge channel distribution data.

[0036] S402: Based on the distribution data of the reflection edge channel, compare the grayscale distribution curves of the dominant direction channel and the adjacent channels, analyze the differences in grayscale distribution trends between adjacent channels, screen the pixel combinations in which grayscale trends change, and obtain the list of interference pixels. The dominant and adjacent channels of each edge pixel are read sequentially, constructing a grayscale sequence of 5 consecutive pixels in both directions. This sequence is then converted into a grayscale trend curve. By sequentially accumulating the grayscale differences between adjacent pixels in each sequence, the overall trend change type is determined. If the dominant direction curve shows a linear upward trend while the adjacent channel curve exhibits a sudden drop or fluctuation (i.e., when the grayscale difference changes from positive to negative with an amplitude exceeding 15), a significant trend difference is identified between the channels. This trend difference combination is recorded, and pixel coordinates are extracted as interference candidates. Further screening is performed if the fluctuation amplitude of the dominant direction grayscale curve is stable at less than 10 at each step. Furthermore, if the grayscale difference fluctuation in adjacent directions exceeds 20, it is determined to be an interfering pixel. For example, if the grayscale value of a pixel in the horizontal direction is 130, 138, 145, 151, 158, and the step increment is stable between 7 and 8, while its grayscale value in the vertical direction is 135, 142, 120, 148, 165, the grayscale difference in the third step is -22, which is a significant fluctuation value, then this pixel is recorded as an interfering pixel. Repeatedly traverse all reflective edge pixels, compare the channel grayscale curve change trend one by one and screen for the existence of interfering fluctuation combinations, record all pixels that meet the abnormal grayscale trend change and their channel combination relationship, and form an interfering pixel identification list.

[0037] S403: Determine the grayscale response distribution of the channels involved in the interference pixel identification list, analyze the proportional relationship of weight changes between the dominant channel and adjacent channels, adjust the weight allocation parameters of the dominant direction and adjacent direction channels, and obtain the directional channel weight configuration. The proportional relationship between the weight changes of the dominant channel and adjacent channels is analyzed using the following formula: ; Calculate the difference weight ratio of the primary and adjacent channels, adjust the weight allocation parameters of the dominant direction and adjacent direction channels, and obtain the directional channel weight configuration, where, This represents the weighting ratio of the differences between primary and secondary channels. This represents the average normalized weight value of the effective pixels in the dominant direction channel. This represents the average normalized weight value of the effective pixels in the adjacent directional channels. This represents the average normalized grayscale response value of the effective pixels in the dominant direction channel. This represents the average normalized grayscale response value of the effective pixels in the adjacent directional channels; Primary and secondary channel difference weight ratio ( The ratio is a normalized quantitative representation of the difference in weight distribution and gray-scale response distribution between the dominant directional channel and the adjacent directional channel. This ratio is achieved by simultaneously examining the differences in average weight values ​​and average gray-scale response values ​​between the two channels, summing them, and adding a constant for normalization. It reflects the relative divergence in feature expression between the two channels; a larger ratio indicates a more significant difference in weight allocation and gray-scale response between the dominant and adjacent channels; a smaller ratio indicates a closer similarity in their performance. This indicator provides a numerical basis for subsequent weight parameter adjustments and directional channel configuration.

[0038] Pixel-level channel features were extracted from the four directional channels of the image. Based on the identified multi-directional gradient response sequences, the horizontal direction was selected as the dominant directional channel and the vertical direction as the adjacent directional channel. The original channel weight values ​​and grayscale response values ​​of four typical pixels in each directional channel were collected. The original weight values ​​of the dominant directional channel were {0.81, 0.86, 0.89, 0.84}, and the original weight values ​​of the adjacent directional channel were {0.64, 0.59, 0.61, 0.62}. The original grayscale response values ​​of the dominant directional channel were {162, 175, 168, 170}, and the original grayscale response values ​​of the adjacent directional channel were {142, 150, 149, 148}. The data were transformed to the [0,1] interval using linear proportional normalization. The denominator of the weight normalization was the difference between the maximum and minimum values, i.e., the dominant direction was... The adjacent direction is The grayscale response normalization is processed in the same way, and the grayscale value range in the dominant direction is... The adjacent direction is The normalized weight values ​​for the dominant direction channel are obtained as {0, 0.625, 1, 0.375}, the normalized weight values ​​for the adjacent direction channel are {1, 0, 0.4, 0.6}, the normalized grayscale response values ​​for the dominant direction are {0, 1, 0.462, 0.615}, and the normalized grayscale response values ​​for the adjacent direction are {0, 1, 0.875, 0.75}. The average values ​​of the four sets of data are calculated respectively. ; ; ; ; Substitute the above values ​​into the formula for calculating the weighted difference ratio: ; ; The total of the numerators is The denominator is Substituting into the calculation, we get: ; This result indicates the difference weighting ratio between primary and secondary channels. Belongs to the interval This interval is defined as the directional channel response consistency segment, and the corresponding structural state is "no significant difference between channels, maintain the original weight configuration"; it represents that the dominant channel and adjacent channels in the current region maintain continuity and stability in feature expression. The result is directly related to the judgment condition of whether to adjust the directional channel weight parameters, so there is no need to change the weight allocation structure. This value is directly used in the weight mapping stage of the subsequent directional channel fusion process. when When defined as a "slightly fluctuating channel difference zone", the dominant channel needs to be proportionally reduced and the adjacent channel needs to be given a compensation factor. like If the channel response is mismatched, it is considered a "channel response mismatch segment" and explicit weight reconstruction and channel priority adjustment operations are performed.

[0039] like Figure 6 As shown, the specific steps for obtaining the skin texture distribution results are as follows: S501: Based on the directional channel weight configuration, analyze the adjusted channel response data, determine the response characteristics of pixel arrangement in each channel, optimize the correspondence between channel direction and original pixel coordinates, calculate the response trend of each channel at the same coordinate point, and obtain the channel fusion response sequence. The system reads the weighted grayscale response value sequence of each pixel across the four directional channels. It iterates through the image point by point according to pixel coordinates, multiplying the response value of each channel with its corresponding channel weight to calculate the weighted response value for that directional channel. This weighted result is then bound to the original pixel coordinates. The consistency of the response intensity between the adjusted channel and the original pixel position is compared for judgment. If the difference in coordinates across multiple channels does not exceed 15, the channel response is considered consistent with the original pixel distribution structure, and the data for each direction is retained. Otherwise, channel response values ​​with a shift exceeding 20 are discarded, and the direction of the retained channel is recorded. For example, at a certain pixel position, the weighted response values ​​for the four directions are 120, 13, and 120 respectively. 2. The vertical response value is 132, and the horizontal response value is 120, which differs by 12, satisfying the consistency condition. Therefore, the responses in both directions are retained. The left diagonal value is 128, which differs from the other directions by 8, and is also retained. The right diagonal value is 125, which differs from the vertical direction by 7, and is also retained. Thus, this pixel retains four channel values. The weighted average of the retained channel response values ​​is taken to form the fused response result. For example, the fused response of this point is (120×0.9+132×1.0+128×1.1+125×1.0) / 4, which is approximately equal to 126.8. Finally, each pixel corresponds to a set of fused response values. After completing the above operations on the entire image, the channel response structure after fusion of all coordinate points is output, forming a channel fused response sequence.

[0040] S502: Based on the channel fusion response sequence, compare the spatial distribution differences with the ambient light reference data, analyze the correlation between the fused grayscale response of each pixel and the changes in ambient light, screen pixels with grayscale trend changes in multiple channels, statistically aggregate spatial features, and obtain the pixel index of the reflection area. Pixel-by-pixel, the fused response value is compared with the corresponding brightness value in the ambient light reference data. If the difference between the response value and the reference brightness value exceeds 25, and the average increase in the response value of the pixel continues to rise in the three directional channels, it is determined that it is significantly affected by ambient light enhancement. This pixel is initially marked as a candidate for reflection enhancement. Then, it is checked whether there are more than four pixels with the same similar fused response trend in its 3×3 neighborhood. If the aggregation criteria are met, the pixel and the center point of the neighborhood are recorded as reflection enhancement pixels. For example, if the fused response value of pixel A is 162 and the corresponding ambient light reference value is 128, the difference is 34. If a threshold is set and the horizontal response is 154, the vertical response is 159, and the left diagonal response is 164, showing an increasing trend, and the five neighboring pixels also show a similar increasing trend, then pixel A and the center position of the neighboring pixels are written into the reflective pixel list. Subsequently, region clustering is performed using this type of pixel as the seed point, and all regions that satisfy spatial connectivity and trend consistency are counted. If the area is greater than 100 pixels, it is defined as a stable reflective region, and its boundary coordinates and number of pixels are recorded. After traversing all pixels to complete the reflective region detection, all pixels that meet the conditions are indexed and output to form a reflective region pixel index set.

[0041] S503: Based on the pixel index of the reflection region, analyze the directional correspondence between the multi-channel fusion response and the skin partition structure, screen the regions that meet the partition texture direction, aggregate the partition direction data, and obtain the skin texture distribution results; The boundary coordinates and center position of each reflection enhancement region are extracted. At the same time, the dominant direction label of the fused response in the corresponding region is called. The number of direction consistency in each small block is counted according to the dominant direction in the region. If the direction consistency ratio is more than 70%, that is, at least 70% of the pixels in the region have the same dominant direction, the direction feature of the region is determined to be stable. Then, the dominant direction is compared with the preset skin partition structure direction template. If the deviation angle between the region's dominant direction and a certain skin partition template is less than 15 degrees, for example, if the region's dominant direction is horizontal and the skin T-zone texture template direction is 0 degrees, the difference between the two is 0 degrees, which meets the matching condition. Then, the region is marked as a T-zone texture candidate region. If the dominant direction of another region is 45 degrees and the corresponding nose wing direction template is 50 degrees, the difference is 5 degrees, which also meets the matching requirement. It is then classified into the nose wing texture region. Subsequently, the texture direction, center position and boundary coordinates of each region are integrated and matched. The texture information of all regions is summarized and the structured data including the direction distribution map, texture direction label and response intensity matrix is ​​output to form the skin texture distribution result.

[0042] like Figure 7 As shown, an AI-based facial skin texture feature extraction system includes: The illumination sensing module is based on an image acquisition device. It records the ambient brightness distribution through a light source sensor, adjusts the shooting angle and exposure parameters, screens effective pixels according to standards, evaluates the uniformity of grayscale pixel brightness, determines the continuous area of ​​pixel distribution, and obtains grayscale spatial distribution data. The gradient extraction module is based on grayscale spatial distribution data, compares the grayscale differences of pixels in multiple directions, analyzes the changing trend using a sliding window, judges the jump coordinates by combining the high contrast feature of skin, identifies the feature display channel, and obtains a multi-directional gradient response sequence. The orientation recognition module is based on multi-directional gradient response sequences to determine the trend of dominant orientation change, compares the grayscale difference between the dominant orientation and each orientation, and combines the neighborhood difference distribution to identify pixels with inconsistent trends, thus obtaining a set of dominant orientation change identifiers. The weight adjustment module analyzes the feature distribution data of the dominant change identifier set, identifies the reflection enhancement edge area, compares the grayscale curves of the dominant and adjacent channels based on the adjacent channel data, identifies interfering pixels and adjusts the weight allocation to obtain the directional channel weight configuration. The texture fusion module maps the fusion response to the original coordinates based on the directional channel weight configuration, screens areas with strong reflection for ambient light data, evaluates structural correspondence, and integrates channel response values ​​to obtain the skin texture distribution results.

[0043] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for extracting facial skin texture features based on AI, characterized in that, The method includes: S1: Based on the image acquisition device, the ambient brightness distribution is recorded through the light source sensor, the shooting angle and exposure parameters are adjusted, effective pixels are screened according to the standard, the uniformity of grayscale pixel brightness is evaluated, the continuous area of ​​pixel distribution is determined, and grayscale spatial distribution data is obtained. S2: Based on the grayscale spatial distribution data, compare the grayscale differences of pixels in multiple directions, analyze the changing trend using a sliding window, determine the jump coordinates by combining the skin high contrast feature law, identify the feature display channel, and obtain a multi-directional gradient response sequence. S3: Based on the multi-directional gradient response sequence, determine the dominant direction change trend, compare the gray level difference between the dominant direction and each direction, and identify pixels with inconsistent trends by combining the neighborhood difference distribution to obtain the dominant direction change identifier set; S4: Analyze the feature distribution data in the set of dominant directional change identifiers, identify the reflection enhancement edge region, compare the grayscale curves of the dominant and adjacent channels based on the adjacent channel data, identify interfering pixels and adjust the weight allocation to obtain the directional channel weight configuration. S5: Based on the directional channel weight configuration, map and fuse the response to the original coordinates, screen for areas with strong reflection in the ambient light data, evaluate the structural correspondence, and integrate the channel response values ​​to obtain the skin texture distribution results.

2. The AI-based facial skin texture feature extraction method according to claim 1, characterized in that, The grayscale spatial distribution data includes a brightness distribution map, pixel continuity and regional uniformity parameters; the multi-directional gradient response sequence includes a direction change data set, jump region information and channel response feature set; the dominant direction change identifier set includes dominant direction labels, trend differentiation markers and feature pixel identifiers; the directional channel weight configuration includes main channel weight parameters, weight allocation index and interference region weight data; and the skin texture distribution result includes a direction distribution map, texture partitioning information and intensity expression matrix.

3. The AI-based facial skin texture feature extraction method according to claim 1, characterized in that, The specific steps for obtaining the grayscale spatial distribution data are as follows: S101: Based on the image acquisition device, analyze the ambient brightness distribution information collected by the light source detection sensor, determine the trend of light change in each area, calculate the brightness continuity between adjacent pixels, and statistically analyze the balance state between pixel areas according to the spatial arrangement change law to obtain spatial brightness distribution data. S102: Based on the spatial brightness distribution data, compare the spatial distribution of pixel grayscale, filter the grayscale data of all pixels in the image frame, judge the validity of pixel information according to the exposure reference standard, filter out pixels that do not meet the reference standard, and obtain a set of qualified grayscale pixels. S103: Based on the set of qualified grayscale pixels, analyze the grayscale distribution uniformity of the pixels, use a sliding window to gradually compare the coherence of grayscale changes in the region, determine the continuous aggregation trend of pixels in the space, optimize the spatial structure of pixel distribution, and obtain grayscale spatial distribution data.

4. The AI-based facial skin texture feature extraction method according to claim 1, characterized in that, The specific steps for obtaining the multi-directional gradient response sequence are as follows: S201: Based on the grayscale spatial distribution data, determine the grayscale level changes of pixels in the horizontal, vertical and diagonal directions, compare the grayscale level changes of adjacent pixels in turn through a sliding window, calculate the continuous change features of pixel sequences in different directions, filter out directional distributions with cumulative change trends, and obtain multi-directional gradient feature groups. S202: Based on the multi-directional gradient feature group, compare the magnitude of gray level changes in each direction, determine the direction in which the pixel's gray level changes are prominent in multiple directions, locate the boundary region according to the pixel's change sequence in each direction, filter the pixel coordinates of gray level changes, and obtain the jump pixel index set. S203: Based on the jump pixel index set, analyze the directional change order of the corresponding pixels, compare the gray level change characteristics of each direction, determine the spatial distribution of gray level changes between channels, integrate the channel characteristics of gray level change regions in different directions, and obtain a multi-directional gradient response sequence.

5. The AI-based facial skin texture feature extraction method according to claim 1, characterized in that, The specific steps for obtaining the direction-dominant change identifier set are as follows: S301: Based on the multi-directional gradient response sequence, determine the gray level change trend of pixels in each directional channel, statistically analyze the gray level change trend of pixels in each directional channel, compare the continuity and offset direction of gray level change in each channel, and obtain the dominant channel sorting information. S302: Based on the dominant channel sorting information, compare the grayscale change amplitude of the dominant direction with that of other directional channels, calculate the cumulative offset of grayscale level differences between channels, determine the coordination of changes between channels in each direction, identify the pixel region where the grayscale distribution difference of the channels is concentrated, and obtain the directional difference feature index. S303: Based on the directional difference feature index, determine the relationship between the dominant direction and the directional distribution of adjacent pixels, analyze the trend of change of the dominant direction and the distribution of differences in surrounding pixels, screen pixels whose directional change trend differs from that of the neighborhood, integrate pixel and dominant direction information, and obtain a set of dominant directional change identifiers.

6. The AI-based facial skin texture feature extraction method according to claim 1, characterized in that, The specific steps for obtaining the directional channel weight configuration are as follows: S401: Based on the dominant direction change identifier set, determine the spatial distribution and directional feature continuity of pixels within the marked area, filter edge pixels that exhibit illumination enhancement in adjacent channels, statistically analyze the spatial combination relationship of pixels, and obtain reflection edge channel distribution data. S402: Based on the reflection edge channel distribution data, compare the grayscale distribution curves of the dominant direction channel and adjacent channels, analyze the differences in grayscale distribution trends between adjacent channels, screen pixel combinations in which grayscale trends change, and obtain an interference pixel identification list. S403: Determine the grayscale response distribution of the channels involved in the interference pixel identification list, analyze the proportional relationship of weight changes between the dominant channel and adjacent channels, adjust the weight allocation parameters of the dominant direction and adjacent direction channels, and obtain the directional channel weight configuration.

7. The AI-based facial skin texture feature extraction method according to claim 1, characterized in that, The specific steps for obtaining the skin texture distribution results are as follows: S501: Based on the directional channel weight configuration, analyze the adjusted channel response data, determine the response characteristics of pixel arrangement in each channel, optimize the correspondence between channel direction and original pixel coordinates, calculate the response trend of each channel at the same coordinate point, and obtain the channel fusion response sequence. S502: Based on the channel fusion response sequence, compare the spatial distribution differences with the ambient light reference data, analyze the correlation between the fusion grayscale response of each pixel and the ambient light change, filter the pixels with grayscale trend changes in multiple channels, statistically analyze the spatial aggregation features, and obtain the pixel index of the reflection area. S503: Based on the pixel index of the reflection area, analyze the directional correspondence between the multi-channel fusion response and the skin partition structure, screen the areas that meet the partition texture direction, aggregate the partition direction data, and obtain the skin texture distribution result.

8. The AI-based facial skin texture feature extraction method according to claim 1, characterized in that, The ambient brightness distribution data refers to the light intensity distribution at various locations in the entire acquisition scene obtained by the light source detection sensor. The effective pixels refer to the set of pixels that meet the conditions after screening for exposure, occlusion, and noise in the acquired original image.

9. The AI-based facial skin texture feature extraction method according to claim 1, characterized in that, The high-contrast skin features refer to the areas in a facial skin image where local grayscale changes and bright-dark contrasts are prominent due to pores, wrinkles, and acne scars. The enhanced reflection edge area refers to the local edge area in a facial image where brightening and specular reflection are caused by skin oil, moisture, or light reflection.

10. An AI-based facial skin texture feature extraction system, the system being used to implement the AI-based facial skin texture feature extraction method as described in any one of claims 1-9, characterized in that, The system includes: The illumination sensing module is based on an image acquisition device. It records the ambient brightness distribution through a light source sensor, adjusts the shooting angle and exposure parameters, screens effective pixels according to standards, evaluates the uniformity of grayscale pixel brightness, determines the continuous area of ​​pixel distribution, and obtains grayscale spatial distribution data. The gradient extraction module compares the grayscale differences of pixels in multiple directions based on the grayscale spatial distribution data, analyzes the changing trend using a sliding window, judges the jump coordinates by combining the high contrast feature of skin, identifies the feature display channel, and obtains a multi-directional gradient response sequence. Based on the multi-directional gradient response sequence, the direction recognition module determines the trend of dominant direction change, compares the grayscale difference between the dominant direction and each direction, and identifies pixels with inconsistent trends by combining the neighborhood difference distribution to obtain a set of dominant direction change identifiers. The weight adjustment module analyzes the feature distribution data of the dominant change identifier set, identifies the reflection enhancement edge region, compares the grayscale curves of the dominant and adjacent channels based on the adjacent channel data, identifies interfering pixels and adjusts the weight allocation to obtain the directional channel weight configuration. The texture fusion module maps and fuses responses to the original coordinates based on the directional channel weight configuration, screens areas with strong reflection for ambient light data, evaluates structural correspondence, and integrates channel response values ​​to obtain skin texture distribution results.

Images