An ai talent skill portrait generation method and system based on multi-modal image recognition

Abstract

The application provides an AI talent skill portrait generation method and system based on multi-modal image recognition, and relates to the field of artificial intelligence.The method comprises the following steps: acquiring multi-modal image data of AI talents in different scenes, and pre-processing the multi-modal image data to obtain pre-processed image data, wherein the multi-modal image data comprises visual modal image data and visual-semantic association modal image data; and positioning a core feature point set related to AI talent skill performance according to the pre-processed image data.Through the collaborative optimization of multiple technical links, the application finally realizes efficient and accurate generation of a skill portrait from multi-modal data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence technology, in particular to an AI talent skill portrait generation method and system based on multi-modal image recognition. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, the ability evaluation and portrait generation of AI talents have become a key link in talent management, post matching and personalized training. Some traditional skill portrait generation methods mainly rely on single modal data, and therefore have the following defects:

[0003] The limitation of single modal data leads to incomplete skill representation. Some traditional methods extract features through visual images or text logs, which are difficult to capture the physical signals of abstract skills, resulting in an evaluation blind area of "seeing but not touching"; cross-modal data alignment is difficult. The actual skill performance of AI talents often involves multi-scene and multi-source data, but different modal data have problems such as timestamp misalignment, light difference and inconsistent spatial scale, which leads to feature correlation fracture and cannot accurately reflect the overall logic of skill activities. SUMMARY

[0004] The technical problem to be solved by the present application is to provide an AI talent skill portrait generation method and system based on multi-modal image recognition, which realizes efficient and accurate generation of skill portraits from multi-modal data through collaborative optimization of multiple technical links.

[0005] To solve the above technical problems, the technical solution of the present application is as follows:

[0006] In a first aspect, an AI talent skill portrait generation method based on multi-modal image recognition is provided, which comprises:

[0007] Step 1: Obtain multi-modal image data of AI talents in different scenes, and pre-process the multi-modal image data to obtain pre-processed image data, wherein the multi-modal image data includes visual modal image data and visual-semantic associated modal image data;

[0008] Step 2: According to the pre-processed image data, locate a core feature point set related to the skill performance of AI talents;

[0009] Step 3: Divide the image into a plurality of Voronoi cells according to the core feature point set, and calculate the feature statistics in each Voronoi cell;

[0010] Step 4: Based on the core feature point set, construct the visibility data of the point set, and calculate the shortest topological connection relationship between the point pairs to form a connected topological structure between the point sets;

[0011] Step 5, according to each Voronoi unit, fuse its corresponding feature statistics and its adjacency relationship in the constructed connected topological structure, calculate the feature weight adjustment value of the Voronoi unit;

[0012] Step 6, extract the initial weight of the feature region of each Voronoi unit in the preprocessed image, and according to the obtained feature weight adjustment value, adaptively optimize the initial weight of the corresponding feature region extracted, and obtain the optimized multi-dimensional feature vector;

[0013] Step 7, preset the AI talent core skill dimension, establish the mapping association between the optimized multi-dimensional feature vector and the preset AI talent core skill dimension, construct a multi-modal skill feature association library, and generate an initial skill portrait of an AI talent.

[0014] Further, in step 1, multi-modal image data of AI talents in different scenes is obtained, and the multi-modal image data is preprocessed, wherein the multi-modal image data includes visual modal image data and visual-semantic association modal image data, including:

[0015] The 3CCD camera and the multi-spectral sensor are used to collect data to obtain visual modal and visual-semantic modal raw data;

[0016] The visual modal and visual-semantic modal raw data are filtered and time-stamped to obtain cross-modal alignment data;

[0017] The cross-modal alignment data is subjected to reflectance correction and standardization processing to obtain processed data;

[0018] The processed data is used to construct a skill label co-occurrence network through a concept graph model, and a code complexity spectral channel is added to enhance the cross-modal features, and obtain the preprocessed image data.

[0019] Further, in step 2, according to the preprocessed image data, the core feature point set related to the skill performance of the AI talent is located, including:

[0020] According to the preprocessed image data, heat data is generated through an attention mechanism architecture, and preliminary feature points are located through non-maximum suppression;

[0021] According to the spatial distance constraint and the feature similarity, the preliminary feature points are clustered to obtain semantic communities;

[0022] According to the fusion of visual feature kernels and semantic association kernels, the feature weights of the visual and semantic communities are obtained;

[0023] The features of visual and semantic communities are fused into visual+semantic joint hash codes, and a core feature point set related to AI talent skill performance is determined through three-stage refinement positioning, wherein the core feature point set includes facial interaction key points, hand operation key points, environmental interaction key points, and feature points representing skill activities.

[0024] Further, in step 3, according to the core feature point set, the image is divided into a plurality of Voronoi cells, and the feature statistics in each Voronoi cell are calculated, including:

[0025] According to the core feature point set, a plane is divided into a plurality of non-overlapping Voronoi cells through a scan line strategy, wherein each cell corresponds to a core feature point, and the distance from any position in the cell to the generated point is less than the distance to other generated points;

[0026] According to the core feature point, a corresponding relationship between the space region and the skill performance type is established, each Voronoi cell is marked with a category attribute, and three types of features are collected in each Voronoi cell to extract multi-modal features, including visual features, semantic features and spectral features;

[0027] The multi-modal features are cross-modal associated and statistical to obtain the feature statistics in each Voronoi cell.

[0028] Further, in step 4, based on the core feature point set, the visibility data of the point set is constructed, and the shortest topological connection relationship between the point pairs is calculated to form a connected topological structure between the point sets, including:

[0029] The visual visibility relationship between the point pairs is established through the spatial distribution coordinates of the core feature point set to construct the visibility data of the point set:

[0030] According to the visibility data of the point set, the unit semantic label weight is optimized;

[0031] According to the unit semantic label weight, the visual-semantic coupling coefficient is taken as the preferred factor to obtain the shortest topological distance matrix between the point pairs, generate the minimum spanning tree T, and obtain the skill association path;

[0032] According to the skill association path, the distance matrix clustering forms a hierarchical structure to form a connected topological structure between the point sets.

[0033] Further, in step 5, according to each Voronoi cell, the feature statistics corresponding thereto and the adjacent relationship in the connected topological structure constructed thereby are fused to calculate the feature weight adjustment value of the Voronoi cell, including:

[0034] According to each Voronoi unit, fuse its corresponding feature statistics and its adjacency relationship in the constructed connected topological structure, obtain a three-level neighborhood unit set;

[0035] Through the neighborhood feature aggregation mechanism, aggregate the three-level neighborhood unit set to obtain a neighborhood enhanced feature vector;

[0036] Through the multi-core SVM fusion of the unit body and the neighborhood enhanced feature vector fusion, obtain a preliminary weight adjustment coefficient;

[0037] Combine the preliminary weight adjustment coefficient with the skill performance to obtain the feature weight adjustment value of the Voronoi unit.

[0038] Further, in step 6, the initial weight of the feature region of each Voronoi unit in the preprocessed image is extracted, and the initial weight of the corresponding feature region is adaptively optimized according to the obtained feature weight adjustment value, to obtain an optimized multi-dimensional feature vector, including:

[0039] Through the Voronoi unit feature region positioning mechanism, the initial weight of the feature region of each Voronoi unit in the preprocessed image is extracted to obtain an initial weight vector;

[0040] Through the feature weight adjustment value mapping mechanism, the feature weight adjustment value is accurately matched with the corresponding unit to obtain a weight set with an adjustment mark;

[0041] Through the topological centrality dynamic correction of the weight set with the adjustment mark, obtain a centrality enhanced weight;

[0042] Through the skill coverage adaptive balance of the centrality enhanced weight, obtain a region balanced weight;

[0043] Through the region balanced weight, reconstruct the weight through the multi-dimensional feature vector to obtain an optimized multi-dimensional feature vector.

[0044] The second aspect is an AI talent skill portrait generation system based on multi-modal image recognition, including:

[0045] An acquisition module is configured to acquire multi-modal image data of an AI talent in different scenarios, and pre-process the multi-modal image data to obtain pre-processed image data, wherein the multi-modal image data includes visual modal image data and visual-semantic associated modal image data;

[0046] A processing module is configured to locate a core feature point set related to the skill performance of the AI talent according to the pre-processed image data;

[0047] The optimization module is configured to divide the image into a plurality of Voronoi cells according to the core feature point set, calculate feature statistics in each Voronoi cell, construct visibility data of the point set based on the core feature point set, calculate the shortest topological connection relationship between point pairs, and form a connected topological structure between the point sets; according to each Voronoi cell, the feature weight adjustment value of the Voronoi cell is calculated by fusing the corresponding feature statistics and the adjacent relationship in the connected topological structure constructed thereby; the initial weight of the feature region of each Voronoi cell in the preprocessed image is extracted, and the initial weight of the corresponding feature region is adaptively optimized according to the obtained feature weight adjustment value, so as to obtain the optimized multi-dimensional feature vector;

[0048] The generation module is configured to preset an AI talent core skill dimension, establish a mapping association between the optimized multi-dimensional feature vector and the preset AI talent core skill dimension, construct a multi-modal skill feature association library, and generate an initial skill portrait of an AI talent.

[0049] In a third aspect, a computing device includes:

[0050] One or more processors;

[0051] A storage device configured to store one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0052] In a fourth aspect, a computer-readable storage medium stores a program, and the program is executed by a processor to implement the method.

[0053] The above-mentioned scheme of the present application at least has the following beneficial effects:

[0054] Firstly, through the cooperative acquisition of 3CCD cameras and multispectral sensors and cross-modal alignment preprocessing, environmental noise, illumination differences and time sequence dislocation problems are effectively eliminated. Secondly, based on the core feature point positioning of attention mechanism and joint hash code, combined with three-level refining screening, sub-pixel level precision is realized, and key features such as facial interaction, hand operation, environmental interaction and skill activity representation are accurately captured, avoiding the missed detection of abstract skills by a single visual mode. Thirdly, through the adaptive division of Voronoi unit and multi-modal feature statistics, the skill performance is quantified as a unit feature vector that does not overlap with each other, breaking through the mechanical restriction of traditional fixed grid and realizing the explicit modeling of the internal logic of skill activities. Based on the connectivity topology structure construction of visibility data and topological centrality, combined with weight adjustment value calculation and multi-dimensional feature vector optimization, noise interference is effectively suppressed, and the weight distribution precision of key skill areas is improved. Finally, through the dynamic rule engine and the persistent storage of multi-modal skill feature association library, adaptive mapping and continuous evolution from features to skill dimensions are realized, and the generated initial skill portrait supports fine-grained skill disassembly, cross-scene ability comparison and bottleneck area visualization. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 is a flowchart of an AI talent skill portrait generation method based on multi-modal image recognition provided by an embodiment of the present application.

[0056] Figure 2 is a schematic diagram of an AI talent skill portrait generation system based on multi-modal image recognition provided by an embodiment of the present application. DETAILED DESCRIPTION

[0057] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0058] As Figure 1 shown, an AI talent skill portrait generation method based on multi-modal image recognition is provided by an embodiment of the present application, which comprises the following steps:

[0059] Step 1, acquiring multi-modal image data of AI talents in different scenes, and preprocessing the multi-modal image data to obtain preprocessed image data, wherein the multi-modal image data includes visual modal image data and visual-semantic associated modal image data;

[0060] Step 2, based on the pre-processed image data, locate a set of core feature points related to AI talent skill performance;

[0061] Step 3, based on the core feature point set, divide the image into a plurality of Voronoi cells, and calculate the feature statistics in each Voronoi cell;

[0062] Step 4, based on the core feature point set, construct the visibility data of the point set, and calculate the shortest topological connection relationship between the point pairs to form the connected topological structure between the point sets;

[0063] Step 5, according to each Voronoi cell, fuse its corresponding feature statistics and the adjacency relationship in the connected topological structure constructed thereby, and calculate the feature weight adjustment value of the Voronoi cell;

[0064] Step 6, extract the initial weight of the feature region of each Voronoi cell in the pre-processed image, and according to the obtained feature weight adjustment value, adaptively optimize the initial weight of the corresponding feature region extracted, to obtain an optimized multi-dimensional feature vector;

[0065] Step 7, preset the core skill dimensions of AI talents, establish a mapping association between the optimized multi-dimensional feature vector and the preset core skill dimensions of AI talents, construct a multi-modal skill feature association library, and generate an initial skill portrait of AI talents.

[0066] In the embodiments of the present application, the technical solutions of the present application significantly improve the precision and depth of AI talent skill evaluation through multi-modal data fusion and spatial topology modeling. Through the fusion of visual and visual-semantic modal data collection and strict cross-modal alignment preprocessing in step 1, combined with the accurate positioning of core feature points based on attention mechanism and joint hash code in step 2, the limitations of single data source are effectively overcome, and the robustness and discriminability of AI talent skill performance feature extraction in complex scenes are significantly enhanced. On this basis, through the adaptive division of Voronoi unit based on core feature points in step 3 and the multi-modal feature statistics in the unit, combined with the shortest topological path constructed in step 4 based on point set visibility data and semantic weight optimization, a hierarchical connected topological structure is formed to depict the spatial correlation and dependency between skill performance regions. This breaks through the mechanical limitations of traditional fixed grid division, and realizes the explicit modeling of the inherent logic of skill activities. Further, through step 5, the feature weight adjustment value is calculated by fusing the unit's own feature statistics and its adjacency relationship in the topological structure, and driving step 6 to adaptively optimize the initial weight based on topological centrality correction and skill coverage rate balance, and finally obtaining a highly optimized multi-dimensional feature vector. This process significantly improves the weight allocation accuracy of key skill regions and effectively suppresses noise interference. Through step 7, an optimized feature vector and a dynamic mapping association library of preset core skill dimensions are established, generating a structured and quantifiable AI talent initial skill portrait.

[0067] In a preferred embodiment of the present application, step 1 can include:

[0068] Step 11, data is collected by 3CCD camera and multi-spectral sensor to obtain visual modal and visual-semantic modal raw data;

[0069] Step 12, filtering and timestamp synchronization are performed on the visual modal and visual-semantic modal raw data to obtain cross-modal alignment data;

[0070] Step 13, reflectance correction and standardization processing are performed on the cross-modal alignment data to obtain processed data;

[0071] Step 14, the processed data is used to construct a skill label co-occurrence network through a concept graph model, and a code complexity spectral channel is added to enhance cross-modal features, and obtain preprocessed image data.

[0072] In the embodiment of the present application, through the cooperative acquisition of 3CCD cameras and multispectral sensors, high-fidelity visual modal data and non-visible spectral data containing semantic information are synchronously acquired, and environmental noise and device difference interference are effectively eliminated by combining cross-modal timestamp alignment and reflectivity correction technology; further, a skill label symbiotic network is constructed by using a concept graph model, and a code complexity spectral channel is innovatively fused to convert abstract programming ability into quantifiable cross-modal features, and finally an enhanced preprocessed image data is generated. The temporal and spatial dislocation problem of multi-source heterogeneous data and the light sensitivity are solved to ensure the robustness of feature extraction; the concept correlation is explicitly modeled through the skill label symbiotic network to break through the semantic sparsity bottleneck of traditional image data; the code complexity spectral channel is enhanced to realize the physical signal conversion of abstract skills such as programming ability, and the representation ability of multi-modal data to AI talent implicit skills is overall significantly improved, especially the objective quantitative analysis basis of the core index of code logic complexity in the development scene is strengthened.

[0073] In a specific embodiment of the present application, the specific steps include:

[0074] Step 11, through the cooperative work of 3CCD cameras and multispectral sensors, visual modal and visual-semantic associated modal original data of professional persons whose professional skills can be observed and analyzed through their multi-modal behavior data (especially visual, visual-semantic associated image data) in specific work scenes while actually operating, applying or researching AI technology in work. The 3CCD camera is responsible for capturing high-resolution visual image data such as programming interface thermal data and code editing trajectory, which can clearly record the operation-intensive area, at the same time, the multispectral sensor covers multiple spectral bands to acquire visual-semantic associated modal data, and a code complexity spectral channel is added, which quantifies the time complexity and space complexity of the algorithm as spectral reflectance, thereby expanding the data dimension and directly reflecting the difficulty and complexity of skill activities.

[0075] Step 12, noise filtering and timestamp alignment are performed on the original data collected in step 11, median filtering is used to process visual image data to remove salt and pepper noise, at the same time, band pass filtering is applied to multispectral signals to retain effective bands and remove irrelevant interference; for semantic data of voice to text, irrelevant content is removed through keyword filtering. Then, through linear interpolation algorithm, timestamp synchronization is realized based on the master clock of the 3CCD camera to ensure that visual operation, document collaboration and voice log are strictly aligned in time sequence (error is controlled within 10ms), for example, code editing operation at a certain time is accurately matched with the corresponding document modification record. This step generates cross-modal aligned data, solves the problems of data noise and time sequence dislocation, and ensures the accuracy of subsequent feature extraction.

[0076] Step 13, based on the cross-modal alignment data of step 12, perform reflectance correction and standardization to eliminate light differences and unify data scales. Among them, the reflectance correction uses a four-step gray scale board as a reference standard, converts the original gray value to true reflectance through a gray value fitting algorithm, and then normalizes the corrected reflectance data to eliminate the dimensional differences of different sensor channels. At the same time, histogram equalization is applied to the visual image to enhance the contrast and highlight the key operation area (such as the code debugging area). This process standardizes the data to a unified scale, for example, the reflectance is mapped to the [0, 1] interval, so that the data of different scenes and sensors is comparable, providing a reliable benchmark for feature statistics.

[0077] Step 14, the data processed in step 13 is enhanced by the concept map model and the new spectral channel to improve the cross-modal correlation. First, based on the preprocessed cross-modal data, the semantic network construction of the concept map model is started: skill tags are extracted from the standardized text logs (code comments, document collaboration content), and the Apriori algorithm is used to analyze the co-occurrence frequency of the tags to identify strong association rules. A weighted directed graph network is constructed: the nodes are skill tags, and the edge weights are conditional probability values (if the probability of "loss function adjustment" appearing when "model parameter adjustment" exists is 0.92). The newly added code complexity spectral channel is activated synchronously: a special detector is added in the 750-900nm near-infrared band of the multispectral sensor, and the Cyclomatic Complexity value of the program code after abstract syntax tree parsing is linearly mapped to the reflectance intensity (complexity 1-10 level corresponds to reflectance 0.1-0.9), and a spectral feature layer aligned with the visual image pixels is generated in real time.

[0078] Then perform cross-modal feature enhancement: superimpose the semantic relationship network output by the concept map to the visual heat map, anchor the semantic tags to the corresponding image area through spatial registration algorithm, and implement feature fusion for label groups with a symbiotic intensity >0.7 (such as "debugging + exception handling"), inject semantic marker color (debugging area superimposes blue channel value +30) in the RGB channel of the corresponding area. At the same time, the spectral reflectance is converted into an HDR heat gradient, and the complexity peak (reflectance >0.8 triggers a red flashing warning) is highlighted in a pulsed manner in the key area of environmental interaction (such as a multi-monitor work area). Finally, a trinity enhanced image is generated: the calibrated heat data retains the original operation density information but enhances the contrast; the semantic hot area marks the skill association cluster with a translucent block; the spectral reflectance is overlaid on the operation interface in the form of contour lines, forming preprocessed image data that can be directly quantified and analyzed.

[0079] In a preferred embodiment of the present application, the above-mentioned step 2 can include:

[0080] Step 21, according to the pre-processed image data, generate heat data through attention mechanism architecture, and locate preliminary feature points through non-maximum suppression;

[0081] Step 22, according to the spatial distance constraint and feature similarity, cluster the preliminary feature points to obtain semantic communities;

[0082] Step 23, according to the fusion of visual feature kernel and semantic association kernel, obtain the feature weight of visual and semantic communities;

[0083] Step 24, fuse the features of visual and semantic communities, and the joint hash code of visual+semantic, and determine the core feature point set related to AI talent skill performance through three-level refinement positioning, wherein the core feature point set includes face interaction key points, hand operation key points, environment interaction key points and feature points representing skill activities.

[0084] In the embodiment of the application, the heat data generated by the attention mechanism realizes the preliminary feature point positioning through non-maximum suppression, and the skill activity semantic community is formed by clustering combined with spatial-semantic double constraints; the joint hash code is generated by innovatively fusing the visual feature kernel and the semantic association kernel to represent the essential properties of the feature points; and finally the four core feature point sets of face interaction, hand operation, environment interaction and skill activity are accurately locked through three-level refinement positioning. The feature points are classified according to the skill activity logic through the collaborative clustering of spatial distance and feature similarity, the explicit behavior and implicit intention are balanced through the visual-semantic weight ratio, and the pure visual method is avoided to miss the abstract skills; the joint hash code fuses the essential properties of multi-modal features, and the three-level refinement effectively filters the environmental noise points; and the overall sub-pixel level core feature point positioning accuracy is realized.

[0085] In a specific embodiment of the application, the specific steps include:

[0086] Step 21, based on the pre-processed multi-modal image data, the saliency regions of visual and semantic modalities are analyzed through the attention mechanism architecture to generate heat data. The heat data filters low response regions through non-maximum suppression technology, retains local peak points as preliminary feature points, such as code high-frequency editing area and document collaboration hotspots, eliminates redundant interference and ensures that the feature points are concentrated in the high-value skill performance area.

[0087] Step 22, the preliminary feature points obtained in step 21 are clustered and analyzed in combination with spatial distance constraints (Euclidean distance threshold) and feature similarity (cosine similarity). Through community discovery algorithm, the dispersed feature points are divided into logically related semantic communities (such as "face interaction community" including eye tracking points and expression recognition points, and "hand operation community" including code editing points and debugging click points), forming a feature group with clear semantic labels.

[0088] Step 23, for the semantic community generated in step 22, the visual feature kernel (SIFT descriptor statistics) and the semantic association kernel (document keyword frequency) are fused to train the multi-kernel support vector machine, and the feature weight of each community is learned. By optimizing the weight ratio of the kernel function (visual kernel weight 60%, semantic kernel weight 40%), the contribution of different communities to skill performance is quantified (such as the "hand operation community" weight is higher than the "environment interaction community").

[0089] Step 24, the weighted community features in step 23, including visual features 64 dimensions + semantic features 64 dimensions, are fused into 128-bit joint hash code, and the core feature points are located through a three-level refinement mechanism:

[0090] Primary refinement: based on Hamming distance to filter similar feature points with hash code, and keep strongly correlated points, such as editing and debugging points in the same code module;

[0091] Intermediate refinement: through topological connectivity verification, such as shortest path constraint between point pairs, to remove isolated points, such as mis-triggered environmental interference points;

[0092] Advanced refinement: combined with skill activity temporal continuity, such as submitting code immediately after debugging, to supplement feature points representing skill evolution, such as code version iteration marker points.

[0093] Finally, a standardized core feature point set is obtained, covering facial interaction key points (focused on expression area), hand operation key points (high-frequency shortcut key area), environmental interaction key points (multi-screen collaboration focus), and skill activity representation points (algorithm optimization hotspots).

[0094] In a preferred embodiment of the present application, step 3 above can include:

[0095] Step 31, according to the core feature point set, the plane is efficiently divided by scanning line strategy, and the image is divided into a plurality of non-overlapping Voronoi cells, wherein each cell corresponds to a core feature point, and the distance from any position in the cell to the generated point is less than the distance to other generated points;

[0096] Step 32, according to the core feature points, the corresponding relationship between the spatial region and the skill performance type is established, each Voronoi cell is marked with a category attribute, and three types of feature collection are performed in each Voronoi cell to extract multi-modal features, including visual features, semantic features and spectral features;

[0097] Step 33, cross-modal association statistics are performed on the multi-modal features to obtain feature statistics in each Voronoi cell.

[0098] In the embodiment of the present application, the plane is efficiently divided by the scan line strategy, the image is divided into "non-overlapping Voronoi cells", and the distance from any position in each cell to the corresponding core feature point is the shortest. In each Voronoi cell, three types of multi-modal features, "visual features, semantic features, and spectral features", are collected synchronously, realizing the all-around capture of skill information. Visual features reflect the physical performance of skill activities, semantic features give abstract semantics to features, and spectral features provide fine features at the physical level; the cooperative collection of the three types of features breaks through the information limitation of a single visual or semantic mode, making the feature representation of each cell more three-dimensional and more comprehensive, covering the "physical performance-semantic connotation-physical property" full dimension of skill activities. Cross-modal correlation statistics of multi-modal features can mine implicit skill rules from isolated modal features. This correlation statistics converts scattered multi-modal features into meaningful skill-related data, so that the feature statistics not only contain the feature attributes of a single mode, but also contain the skill essence reflected by multi-modal cooperation. The scan line strategy is used to realize efficient division of Voronoi cells, which significantly improves the processing efficiency while ensuring the accuracy of spatial division.

[0099] In a specific embodiment of the present application, the specific steps include:

[0100] Step 31, based on the spatial distribution of the core feature point set, the image plane is dynamically divided by the scan line strategy: the improved Fortune algorithm is used to perform plane scanning calculation with the core feature points as the generating points, and efficient division is completed within the time complexity. Each Voronoi cell strictly satisfies the spatial characteristics that the distance from any position in the cell to its corresponding generating point is less than the distance to other generating points, forming a non-overlapping cell grid (such as the "algorithm design cell" covering the high-frequency code editing area and the "document collaboration cell" covering the semantic hot area).

[0101] Step 32, based on the divided Voronoi cells, a space-skill mapping relationship is established: first, according to the semantic type of the core feature points (the clustering result of step 22), the cell is labeled with a category attribute ("face interaction cell" is labeled as an attention analysis area, and "hand operation cell" is labeled as an efficiency monitoring area); then, three types of multi-modal features are collected synchronously in each cell, including:

[0102] Visual features: extract operation heat value (click frequency) and trajectory density (mouse movement path complexity);

[0103] Semantic features: count the frequency of document collaboration keywords (such as "optimize" and "debug");

[0104] Spectral features: quantify the code complexity spectrum reflection (high time complexity corresponds to high near-infrared reflectance).

[0105] Step 33, fusion analysis of intra-unit isomerism features: using canonical correlation analysis (CCA) to calculate the correlation coefficient of visual-semantic features (such as the correlation strength of heat value and keyword frequency), and fusing the spectral reflectance and operation density by weighted average (high complexity area is given higher weight); finally output the feature vector of each unit, (such as [heat value: 0.85, keyword frequency: 0.72, spectral reflectance: 0.91]) and cross-modal correlation matrix, quantifying the skill performance intensity and modality consistency, eliminating redundant information, and reducing the error rate by 15%.

[0106] In a preferred embodiment of the present application, step 4 can include:

[0107] Step 41, the visual visibility relationship between point pairs is established by the spatial distribution coordinates of the core feature point set, and the visibility data of the point set is constructed:

[0108] Step 42, according to the visibility data of the point set, the unit semantic label weight is optimized;

[0109] Step 43, according to the unit semantic label weight, the visual-semantic coupling coefficient is taken as the preferred factor, the shortest topological distance matrix between point pairs is obtained, the minimum spanning tree T is generated, and the skill association path is obtained;

[0110] Step 44, according to the skill association path, the hierarchical structure is formed by distance matrix clustering, and the connected topological structure between point sets is formed.

[0111] In the embodiment of the present application, the spatial distribution coordinates of the core feature point set are used to establish a "visual visibility relationship", which directly quantifies the physical interaction possibility between feature points. Through the weight adjustment strategy of "similar skill area attenuation, cross-skill area enhancement, and bottleneck area significant enhancement", the redundant information of similar skill areas is suppressed, the feature is avoided from being submerged due to excessive concentration of weight, the cross-skill correlation area is highlighted, the feature signal of multi-skill cooperation ability is strengthened, the correlation weight of the skill bottleneck area is highlighted, and the skill short board is more easily identified in the topological structure; the shortest topological distance matrix is calculated by using the "visual-semantic coupling coefficient" as the preferred factor, and the minimum spanning tree (skill correlation path) is generated, realizing two-layer core optimization, preferentially reserving the path with high "visual feature matching degree + close semantic correlation", filtering the noise path without semantic correlation, and the minimum spanning tree structure ensures that all core feature points are covered by the least correlation edge, extracts the key dependency chain in the skill activity, avoids the skill logic confusion caused by redundant correlation, and makes the skill correlation relationship more focused and easier to understand. The distance matrix clustering forms a "hierarchical topological structure", which divides the skill feature points into different levels according to the correlation tightness, and realizes the structured expression of the skill correlation relationship: the upper structure reflects the direct correlation of the core skill; the lower structure embodies the indirect support of the auxiliary skill or environmental factor; this hierarchical structure directly presents the skill system framework of the AI talent.

[0112] In a specific embodiment of the present application, the specific steps include:

[0113] Step 41, based on the spatial coordinate distribution of the core feature point set, the visual visibility relationship between the point pairs is calculated by the line-of-sight projection algorithm: taking each feature point as the observation center, scanning the straight line visibility (no obstacle blocking is considered as visible) between it and the surrounding points, generating a visibility binary matrix (1 for visible points, otherwise 0) between the point set, and marking the skill area attributes (such as "algorithm design area" and "debug bottleneck area") in this process, giving the same skill area a basic weight of 1.0, the cross-skill area a weight of 0.5, and the skill bottleneck area a weight of 0.2, to form an initial visibility data set.

[0114] Step 42, based on the visibility data, the unit semantic label is optimized by the weight attenuation-enhancement mechanism, specifically including:

[0115] Similar skill area: if the visible point pairs belong to the same semantic label (such as "code debugging area"), the weight is attenuated by 30% (weight x 0.7), and the redundant correlation is weakened;

[0116] Cross-skill area: if the visible point pairs belong to different semantic labels (such as "debugging area" and "document collaboration area"), the weight is enhanced by 50% (weight x 1.5), and the skill cross value is strengthened;

[0117] Skill bottleneck area: associate the edge weight of high complexity area (such as time complexity algorithm module) with 200% (weight x 3.0) to highlight the key path.

[0118] Step 43, with the optimized semantic weight as the edge weight, the visual-semantic coupling coefficient (thermal value and keyword frequency correlation coefficient) as the path optimization factor, and the shortest topological distance between points calculated by Dijkstra algorithm, the process includes:

[0119] First, construct a weighted directed graph (node = feature point, edge weight = semantic weight x coupling coefficient);

[0120] Second, generate a global shortest distance matrix with the minimum path cost as the optimization objective;

[0121] Based on the distance matrix, extract the minimum spanning tree T, retain the strong correlation edges (such as "debug to deployment" path), and eliminate the weak connection edges (weight < 0.3) to obtain the skill backbone association path (such as core algorithm optimization chain), and form the logical skeleton of cross-regional skill cooperation.

[0122] Step 44, based on the distance matrix of skill association path, generate hierarchical topological structure by hierarchical clustering algorithm, including:

[0123] Primary clustering: form local skill clusters with Manhattan distance ≤ 2 as cluster centers;

[0124] Intermediate aggregation: merge primary clusters with semantic label similarity ≥ 80% to construct skill modules (such as "algorithm optimization module");

[0125] Global hierarchical: generate tree-like hierarchy according to the topological distance between modules (root node is core skill module, leaf node is sub-function cluster), finally output hierarchical connected topological structure, and intuitively present skill association strength and hierarchical dependence (such as "innovation design layer to engineering implementation layer to debugging support layer").

[0126] In a preferred embodiment of the present application, the above step 5 can include:

[0127] Step 51, according to each Voronoi unit, fuse its corresponding feature statistics and its adjacency relationship in the connected topological structure to obtain a three-level neighborhood unit set;

[0128] Step 52, aggregate the three-level neighborhood unit set through neighborhood feature aggregation mechanism to obtain neighborhood enhanced feature vector;

[0129] Step 53, fuse the preliminary weight adjustment coefficient by fusing the unit ontology and neighborhood enhanced feature vector through multi-core SVM;

[0130] Step 54, combine the preliminary weight adjustment coefficient with the skill performance to obtain the characteristic weight adjustment value of the Voronoi cell.

[0131] In an embodiment of the present application, by constructing a "three-level neighborhood cell set", a single Voronoi cell is associated with its surrounding neighborhood cells of different ranges, breaking the isolated analysis mode of cell characteristics. This design can capture the hierarchical relationship of "local cell-neighborhood correlation-global scene" in skill performance, for example, the hand operation cell needs to be combined with the face concentration cell and the context of the environmental tool cell. The neighborhood feature aggregation mechanism integrates the visual, semantic, spectral and other multi-modal features of the three-level neighborhood cells, so that the feature of a single cell expands from "its own basic feature" to "enhanced feature with integrated neighborhood correlation information"; by fusing "cell intrinsic feature" and "neighborhood enhanced feature" through multi-kernel SVM, the advantages of multi-kernel model in fitting complex multi-modal features are fully utilized: on the one hand, the core feature weight of the cell itself is retained, and on the other hand, the noise interference of single feature (such as feature deviation caused by local image blur) is corrected through neighborhood feature. Combine the preliminary weight adjustment coefficient with the "skill performance" directly, so that the final feature weight adjustment value is no longer dependent on pure data-driven calculation. Through hierarchical neighborhood correlation analysis and multi-model fusion, the "core degree" and "correlation degree" of different Voronoi cells in skill performance can be accurately identified. This differential adjustment mechanism lays a reasonable weight foundation for the optimization of subsequent multi-dimensional feature vectors, and improves the accuracy of skill portrait.

[0132] In a specific embodiment of the present application, the specific steps include:

[0133] Step 51, based on the connected topology structure generated in step 4, extract the three-level neighborhood relationship of each Voronoi cell, wherein the first-level neighborhood is the directly topologically adjacent cell (the directly connected cell in the minimum spanning tree T); the second-level neighborhood is the indirectly connected cell (the cell with a topological distance ≤2); and the third-level neighborhood is the same type of semantic label cell (the cell with the same skill type across regions). Through this process, the cell feature statistics and topological adjacency relationship are integrated to form a neighborhood cell set (for example, the neighborhood of a certain debugging cell includes the directly associated algorithm cell, the indirectly associated deployment cell and the same type of debugging cell).

[0134] Step 52, integrate the three-level neighborhood information through the neighborhood feature aggregation mechanism, including:

[0135] Aggregation strategy: the first-level neighborhood feature is directly weighted average (weight 0.6), the second-level neighborhood feature is weighted inversely proportional to the topological distance (the weight decays by 30% for every 1 unit increase in distance), and the third-level neighborhood feature takes the maximum value of the same type of label (emphasizes common features).

[0136] Generate neighborhood enhanced feature vector (such as the enhanced vector of a certain unit = [body heat value 0.8, neighborhood average debugging frequency 0.75, maximum complexity of the same kind 0.9]).

[0137] Step 53, visual kernel (RBF kernel) processing operation heat value and other continuous features, semantic kernel (linear kernel) processing keyword frequency and other discrete features. Through the kernel weight proportioning (visual kernel 70% + semantic kernel 30%) training SVM model, get the preliminary weight adjustment coefficient (such as the coefficient of a certain algorithm unit 1.2, indicating that its skill contribution needs to be adjusted by 20%).

[0138] Step 54, associate the preliminary weight adjustment coefficient with the skill performance measured data, and perform dynamic calibration: if the actual skill performance of the unit (such as debugging efficiency) is higher than the predicted value, then the coefficient x 1.3 (positive reinforcement); if it is lower than the predicted value, then the coefficient x 0.7 (negative correction). Then, the bottleneck is perceived, and the skill bottleneck area (spectrum reflection amount > 0.9) is additionally given a weight gain of 200% (coefficient x 3.0), focusing on key problem points. Finally, the feature weight adjustment value (such as the adjustment value of a certain high complexity unit = 1.2 x 3.0 = 3.6) is obtained, realizing the dynamic matching of "feature importance-skill performance".

[0139] In a preferred embodiment of the present application, the above step 6 can include:

[0140] Step 61, through the Voronoi unit feature area positioning mechanism, the initial weight of the feature area of each Voronoi unit in the preprocessed image is extracted, and the initial weight vector is obtained;

[0141] Step 62, through the feature weight adjustment value mapping mechanism, the feature weight adjustment value is accurately matched with the corresponding unit to obtain the weight set with adjustment mark;

[0142] Step 63, through the topological centrality dynamic correction of the weight set with adjustment mark, the centrality enhanced weight is obtained;

[0143] Step 64, through the skill coverage rate adaptive balance centrality enhanced weight, the area balanced weight is obtained;

[0144] Step 65, through the area balanced weight, the multi-dimensional feature vector is reconstructed by the multi-dimensional feature vector.

[0145] In the embodiment of the present application, the initial weight vector is extracted by the "Voronoi cell feature region positioning mechanism", which ensures that the initial weight is strictly related to the feature region of each Voronoi cell in the preprocessed image. This extraction method based on spatial positioning avoids the mismatch between weight and feature region, and ensures the spatial accuracy of the initial weight from the source. The "feature weight adjustment value mapping mechanism" is used to realize the accurate matching of the adjustment value and the corresponding cell, and to form a weight set with adjustment marks. The "feature weight adjustment value" is accurately bound to the corresponding Voronoi cell. The weight adjustment of each cell is based on its exclusive feature statistics and topological correlation analysis. The "topological centrality dynamic correction" mechanism is introduced to optimize the weight set with marks and generate centrality enhanced weights. Topological centrality can identify the Voronoi cells that are in the core position in the skill correlation network. By dynamically improving the weight of these core cells, the signal strength of key skill features is strengthened. At the same time, the weight of secondary cells at the topological edge is weakened to avoid the submersion of secondary features to the core skill signal. This correction makes the weight distribution more consistent with the core logic of skill activities, highlighting the core skill advantages of talents. Through the "skill coverage adaptive balance" mechanism, the regional balanced weight is optimized, solving the problem of "core cell weight being too high, causing the secondary but necessary skill features to be ignored". Finally, the "multi-dimensional feature vector reconstruction" generates an optimized multi-dimensional feature vector, which integrates the weight information optimized in the early stage into a feature carrier with compact structure and reasonable dimensions. The reconstruction process not only removes redundant features, but also strengthens the discriminability of the feature vector through weight optimization.

[0146] In a specific embodiment of the present application, the specific steps include:

[0147] Step 61, scan the preprocessed image by the "Voronoi cell feature region positioning mechanism" to extract the initial weight of each cell feature region: according to the strong correlation between the cell space boundary and the feature intensity, based on the feature statistics (such as heat value, spectral reflectance) of step 3, combined with the cell semantic label (such as "debugging area" "algorithm area") to generate an initial weight vector (such as [debugging area weight: 0.7, algorithm area weight: 0.8]).

[0148] Step 62, use the "feature weight adjustment value mapping mechanism" to accurately match the adjustment value (such as debugging area adjustment value +0.3) output by step 5 with the corresponding cell: realize fast matching through cell ID hash coding, add adjustment marks to the weight vector (such as [debugging area: 0.7 to +0.3, algorithm area: 0.8 to -0.1]), and generate a weight set with dynamic marks.

[0149] Step 63, based on the connected topological structure generated in step 4, correct the weight set through topological centrality analysis, which specifically includes:

[0150] Core unit recognition: the computing unit identifies the intermediate centrality of the topology network (e.g., units that frequently appear on the shortest path);

[0151] Weight reinforcement: the weight of high centrality units (e.g., core debugging nodes) is increased by 30% (marked value x 1.3), and the weight of edge units is decreased by 20% (marked value x 0.8).

[0152] Get centrality-enhanced weight (e.g., core debugging area weight from 1.0 to 1.3), highlighting the global influence of key skill paths.

[0153] Step 64, introduce skill coverage adaptive balance mechanism, perform coverage calculation: count the coverage proportion of the same type of skill unit in the topology network (e.g., "debugging type" unit accounts for 25%); weight balance: increase the weight of low coverage skill units (<20%) by 40%, and decrease the weight of high coverage units (>50%) by 10%, to eliminate regional bias, and finally get the regional balanced weight.

[0154] Step 65, integrate regional balanced weight through multi-dimensional feature vector reconstruction algorithm, the specific process includes:

[0155] First, dimension mapping, map the weight value to the preset skill dimension (e.g., innovation, engineering);

[0156] Second, vector synthesis: aggregate weights according to skill dimensions (e.g., all "innovation design" unit weight mean to innovation dimension);

[0157] Finally, normalized output: generate the optimized multi-dimensional feature vector (e.g., [innovation: 0.85, engineering: 0.78, collaboration: 0.68]), realize the dimension conversion from local weight to global skill vector.

[0158] In a preferred embodiment of the present application, the above step 7, the preset AI talent core skill dimension, establishes the mapping relationship between the optimized multi-dimensional feature vector and the preset AI talent core skill dimension, constructs a multi-modal skill feature association library, to generate the initial skill portrait of AI talent, which can include:

[0159] Step 71, according to the historical data, preset the core skill dimension framework, get the preset AI talent core skill dimension library;

[0160] Step 72, according to the optimized multi-dimensional feature vector and the preset AI talent core skill dimension library, construct a dynamic rule mapping engine;

[0161] Step 73, according to the dynamic rule mapping engine, construct a structured association matrix, and process the structured association matrix to generate a multi-modal skill feature association library;

[0162] Step 74, according to the multi-modal skill feature association library, the skill evaluation result is calculated in the three-dimensional polar coordinate system to generate the initial skill portrait of AI talents.

[0163] In an embodiment of the present application, based on the skill dimension library preset according to historical data and the optimized multi-dimensional feature vector, adaptive matching is realized through a dynamic rule engine, overcoming the defect that traditional static evaluation models are difficult to adapt to technical iteration. The engine analyzes the correlation strength of the feature vector and the skill dimension in real time, dynamically adjusts the weight combining with industry evolution data, ensures that the portrait always reflects the latest skill demand of the market; the mapping relationship is constructed as a structured association matrix, quantifying the causal chain of multi-modal features and skill dimensions, the matrix supports traceability analysis, making the talent capability bottleneck visualized; the skill association library data is integrated in the three-dimensional polar coordinate system to generate an interactive initial skill portrait. This design breaks through the plane limitation of traditional radar charts, intuitively presents the three-dimensional structure and dynamic trajectory of the ability, and supports cross-scene comparison.

[0164] In a specific embodiment of the present application, the specific steps include:

[0165] Step 71, extract historical evaluation records (including project performance, skill certification score) from enterprise talent database, and synchronously crawl structured data of AIGC and other authoritative frameworks. After five-stage cleaning process: delete isolated labels with frequency <5% (such as "quantum computing optimization"), unify naming rules ("PyTorch debugging to deep learning framework debugging"), and map to three-layer classification system (technology layer / application layer / collaboration layer). Identify skill association groups through density clustering algorithm, generate atomic ability unit tree structure: each unit contains unique code (such as TECH ALGO 003), standardized name (limited to 20 characters), basic weight (0.5-1.0 initial value) and dynamic factor (default 1.0). The dynamic mechanism grabs GitHub trend library every 6 hours, calculates the heat growth rate of skill labels (monthly new star number / total star number × search index), when the growth rate is >10% for 3 consecutive days, the dynamic factor is increased by 0.1 and the weight recalculation is triggered (new weight = basic weight × dynamic factor), for example, the "large model fine-tuning" unit automatically increases from 0.8 to 0.92 due to the surge in demand. Finally, output the core skill dimension library with version number, support filtering and exporting according to technical direction.

[0166] Step 72, compile skill units into Drools rules: atomic rules consist of condition group (feature threshold judgment) and action group (activate skill unit + confidence assignment), for example, rule R023 is defined as "when [code complexity > 0.7] and [debug frequency > 0.5], activate TECH_CODE_001 unit and assign a weight of 1.3". The engine receives the feature vector and initiates a three-level matching: the first layer quickly filters out rules that do not meet the threshold (such as skipping debug frequency 0.3), the middle layer accurately matches rules that meet all conditions, and the top layer aggregates multiple rule outputs (same unit weight accumulation). The attention weight module perceives the scene in real time: the code complexity spectrum feature weight in the programming scene is x1.8, and the facial expression feature in the collaboration scene is x1.5. Conflict resolution adopts a double strategy: rules triggered by spectrum features automatically obtain the highest priority, and the rest of the conflict is selected by a Q-learning model with a historical return rate > 85% action path. When a new skill dimension (such as AI ethics) is added, the engine automatically loads the precompiled rule package and updates the weight distribution table.

[0167] Step 73, create a sparse matrix of m rows (skill units sorted by level) and n columns (feature dimensions sorted by collection order). The correlation strength calculation combines semantic and topological double factors: 60% weight comes from the cosine similarity between skill description word vector and feature semantic vector (encoded by BERT model), and 40% weight comes from the shortest path coefficient in the topological network (calculated by Dijkstra algorithm). The real-time conflict detection engine scans the matrix elements, when a feature is simultaneously associated with mutually exclusive skills (such as "algorithm optimization" and "rapid prototyping development"), the entropy repair process is started: query the mutual exclusion relationship table of the knowledge graph, calculate the correlation dispersion of features and skills, and implement a 20% weight reduction for correlations with dispersion > 0.7. The matrix uses a hierarchical storage architecture: the basic layer stores full data in CSR format, and the application layer generates sub-matrices according to the scene (200x100 subset for interview evaluation). The incremental update mechanism monitors the weight changes of skill units, and when the fluctuation > 15%, it locates the affected row number, recalculates the values of all columns in that row, and records the version snapshot.

[0168] Step 74, the space mapping engine performs a three-axis conversion: the radial axis (R axis) aggregates skill unit weights, and the innovation type skill is weighted x 1.2 and then the sigmoid normalized output is 0-1 value; the azimuth angle (θ axis) is distributed according to the proportion of technical ability (0°-90°) and cooperation ability (90°-180°) (technical proportion = total weight of technical skills / total weight of the whole x 90°); the polar angle (φ axis) is compared with the current and historical quarterly data, and when the skill strength increases by more than 10%, it is marked +35° upward, and when it decreases by more than 5%, it is marked -25° downward. The visualization engine connects the coordinate points of the last four consecutive quarters through B-spline interpolation to generate an ability evolution curve, and the line width is inversely proportional to the skill stability variance (line width 3px when variance <0.1). The automatic diagnosis module scans the skill points with R<0.4, traces the weak feature column (correlation strength <0.3) through the correlation matrix, and generates a root cause report such as: "Ethical risk identification (R=0.38) is caused by the 'compliance' keyword frequency 0.21 (threshold 0.5) in the design document". Finally, an interactive portrait with space-time coordinates is output, supporting drag and rotate the viewing angle, hover to view skill details, and click on the short board to jump to the original feature data source.

[0169] Embodiments of the present application also provide an AI talent skill portrait generation system based on multi-modal image recognition, comprising:

[0170] The acquisition module is configured to acquire multi-modal image data of AI talents in different scenarios, and pre-process the multi-modal image data to obtain pre-processed image data, wherein the multi-modal image data includes visual modal image data and visual-semantic association modal image data.

[0171] The processing module is configured to locate a core feature point set related to the skill performance of the AI talent according to the pre-processed image data.

[0172] The optimization module is configured to divide the image into a plurality of Voronoi cells based on the core feature point set, calculate feature statistics in each Voronoi cell, construct visibility data of the point set based on the core feature point set, calculate the shortest topological connection relationship between point pairs, and form a connected topological structure between the point sets; calculate the feature weight adjustment value of the Voronoi cell according to each Voronoi cell, fuse the corresponding feature statistics and the adjacent relationship in the connected topological structure constructed thereby, and extract the initial weight of the feature region of each Voronoi cell in the pre-processed image, and adaptively optimize the initial weight of the corresponding feature region extracted according to the obtained feature weight adjustment value to obtain an optimized multi-dimensional feature vector.

[0173] The generating module is configured to preset an AI talent core skill dimension, establish a mapping association between the optimized multi-dimensional feature vector and the preset AI talent core skill dimension, construct a multi-modal skill feature association library, and generate an initial skill portrait of the AI talent.

[0174] It should be noted that the system corresponds to the above method, and all implementation manners in the above method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0175] The above is the preferred embodiment of the present application. It should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should also be considered within the scope of protection of the present application.

Claims

1. An AI talent skill portrait generation method based on multi-modal image recognition, characterized in that, The method comprises: Step 1, acquiring multi-modal image data of AI talents in different scenes, and pre-processing the multi-modal image data to obtain pre-processed image data, wherein the multi-modal image data comprises visual modal image data and visual-semantic association modal image data; Step 2, locating a core feature point set related to the skill performance of AI talents according to the pre-processed image data; Step 3, dividing the image into a plurality of Voronoi cells according to the core feature point set, and calculating feature statistics in each Voronoi cell; Step 4, based on the core feature point set, constructing the visibility data of the point set, and calculating the shortest topological connection relationship between the points to form a connected topological structure between the points; Step 5, according to each Voronoi cell, fusing its corresponding feature statistics and the adjacency relationship in the connected topological structure constructed thereby, and calculating a feature weight adjustment value of the Voronoi cell; Step 6, extracting the initial weight of the feature region of each Voronoi cell in the pre-processed image, and adaptively optimizing the initial weight of the corresponding feature region according to the obtained feature weight adjustment value to obtain an optimized multi-dimensional feature vector; Step 7, presetting the core skill dimension of AI talents, establishing a mapping relationship between the optimized multi-dimensional feature vector and the preset core skill dimension of AI talents, constructing a multi-modal skill feature association library, and generating an initial skill portrait of AI talents. 2.The AI talent skill portrait generation method based on multi-modal image recognition according to claim 1, characterized in that, In step 1, the multi-modal image data of AI talents in different scenes is acquired, and the multi-modal image data is pre-processed, wherein the multi-modal image data comprises visual modal image data and visual-semantic association modal image data, which comprises: The 3CCD camera and the multi-spectral sensor are used to collect data to obtain visual modal and visual-semantic modal raw data; The visual modal and visual-semantic modal raw data are filtered and time-stamped to obtain cross-modal alignment data; The cross-modal alignment data are subjected to reflectivity correction and standardization processing to obtain processed data; The processed data are used to construct a skill label co-occurrence network through a concept graph model, and a code complexity spectral channel is added to enhance the cross-modal features, thereby obtaining pre-processed image data. 3.The AI talent skill portrait generation method based on multi-modal image recognition according to claim 2, characterized in that, In step 2, the core feature point set related to the skill performance of AI talents is located according to the pre-processed image data, which comprises: According to the pre-processed image data, heat data is generated through an attention mechanism architecture, and preliminary feature points are located through non-maximum suppression; The preliminary feature points are clustered according to spatial distance constraints and feature similarity to obtain semantic communities; According to the fusion of visual feature kernel and semantic association kernel, the feature weights of visual and semantic communities are obtained; The features of visual and semantic communities are fused with visual+semantic joint hash codes, and through three-stage refinement positioning, the core feature point set related to the skill performance of AI talents is determined, wherein the core feature point set comprises facial interaction key points, hand operation key points, environmental interaction key points, and feature points representing skill activities. 4.The AI talent skill portrait generation method based on multi-modal image recognition according to claim 3, characterized in that, In step 3, according to the core feature point set, the image is divided into multiple Voronoi cells, and the feature statistics in each Voronoi cell are calculated, including: According to the core feature point set, the image is divided into multiple Voronoi cells by scanning line strategy, and each Voronoi cell corresponds to a core feature point, and the distance from any position in the cell to the generated point is less than the distance to other generated points; According to the core feature point, the corresponding relationship between the space region and the skill performance type is established, each Voronoi cell is marked with a category attribute, and three types of features are collected in each Voronoi cell to extract multi-modal features, including visual features, semantic features and spectral features; The multi-modal features are statistically correlated across modalities to obtain feature statistics in each Voronoi cell. 5.The AI talent skill portrait generation method based on multi-modal image recognition according to claim 4, characterized in that, In step 4, based on the core feature point set, the visibility data of the point set is constructed, and the shortest topological connection relationship between the points is calculated to form a connected topological structure between the points, including: The visual visibility relationship between the points is established based on the spatial distribution coordinates of the core feature point set, and the visibility data of the point set is constructed: According to the visibility data of the point set, the unit semantic label weight is optimized; According to the unit semantic label weight, the visual-semantic coupling coefficient is taken as the preferred factor to obtain the shortest topological distance matrix between the points, generate the minimum spanning tree T, and obtain the skill association path; According to the skill association path, the distance matrix clustering forms a hierarchical structure to form a connected topological structure between the points. 6.The AI talent skill portrait generation method based on multi-modal image recognition according to claim 5, characterized in that, In step 5, according to each Voronoi cell, the corresponding feature statistics and the adjacency relationship in the connected topological structure constructed by it are fused to calculate the feature weight adjustment value of the Voronoi cell, including: According to each Voronoi cell, the corresponding feature statistics and the adjacency relationship in the connected topological structure constructed by it are fused to obtain a three-level neighborhood cell set; The three-level neighborhood cell set is aggregated by a neighborhood feature aggregation mechanism to obtain a neighborhood enhanced feature vector; The preliminary weight adjustment coefficient is obtained by fusing the unit body and the neighborhood enhanced feature vector through multi-core SVM; The feature weight adjustment value of the Voronoi cell is obtained by combining the preliminary weight adjustment coefficient with the skill performance. 7.The AI talent skill portrait generation method based on multi-modal image recognition according to claim 6, characterized in that, In step 6, the initial weight of the feature region of each Voronoi cell in the preprocessed image is extracted, and the initial weight of the corresponding feature region is adaptively optimized according to the obtained feature weight adjustment value to obtain an optimized multi-dimensional feature vector, including: The initial weight of the feature region of each Voronoi cell in the preprocessed image is extracted by a Voronoi cell feature region positioning mechanism to obtain an initial weight vector; The feature weight adjustment value is accurately matched with the corresponding cell through a feature weight adjustment value mapping mechanism to obtain a weight set with adjustment marks; The weight set with adjustment marks is dynamically corrected by topological centrality to obtain a centrality enhanced weight; The region balanced weight is obtained by adaptively balancing the centrality enhanced weight. The multi-dimensional feature vector is obtained by balancing the weights of regions and reconstructing the weights by a multi-dimensional feature vector.

8. An AI talent skill portrait generation system based on multi-modal image recognition, the system implements the method of any one of claims 1 to 7, characterized in that, The method comprises the following steps: An acquisition module is configured to acquire multi-modal image data of AI talents in different scenarios, and to pre-process the multi-modal image data to obtain pre-processed image data, wherein the multi-modal image data comprises visual modal image data and visual-semantic association modal image data. A processing module is configured to locate a core feature point set related to the skill performance of AI talents according to the pre-processed image data. An optimization module is configured to divide an image into a plurality of Voronoi cells according to the core feature point set, to calculate feature statistics in each Voronoi cell, to construct visibility data of the point set based on the core feature point set, to calculate the shortest topological connection relationship between point pairs to form a connected topological structure between the point sets, to calculate a feature weight adjustment value of each Voronoi cell according to the corresponding feature statistics and the adjacent relationship in the connected topological structure constructed thereby, to extract an initial weight of a feature region of each Voronoi cell in the pre-processed image, and to adaptively optimize the initial weight of the corresponding feature region according to the obtained feature weight adjustment value to obtain an optimized multi-dimensional feature vector. A generation module is configured to preset a core skill dimension of AI talents, to establish a mapping relationship between the optimized multi-dimensional feature vector and the preset core skill dimension of AI talents, to construct a multi-modal skill feature association library, and to generate an initial skill portrait of AI talents.

9. A computing device, comprising: The method comprises the following steps: One or more processors; A storage device is configured to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a program which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images