An intelligent visual analysis method, device and medium based on double principal component analysis and YOLO model

Abstract

The application discloses an intelligent visual analysis method and device based on double principal component analysis and a YOLO model and a medium. The method comprises the following steps: performing first PCA dimension reduction on original visual data to obtain first feature representation after denoising; inputting the first feature representation into a YOLO model to perform real-time target detection and obtain a structured result; constructing a feature matrix based on the result and performing second PCA to mine potential patterns and output an interpretable analysis result. Through the closed-loop process of 'dimension reduction-detection-interpretation', the application improves the processing speed and detection accuracy, solves the problem that the result of a deep learning model is not interpretable, realizes the leap from perception to cognition, and can be widely applied to the fields of sports analysis, intelligent traffic, medical imaging and industrial quality inspection.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and machine learning technology. Specifically, it relates to an intelligent visual analysis method, system, and related equipment that integrates dual principal component analysis (PCA) and YOLO (You Only Look Once) target detection models, and is particularly suitable for visual analysis scenarios with high dynamics, multiple targets, and the requirement for interpretable results. Background Technology

[0002] In complex visual scenarios such as sports event analysis, intelligent traffic monitoring, medical image diagnosis and industrial visual quality inspection, real-time processing of massive high-dimensional data and extraction of understandable patterns faces three core challenges: (1) The original visual data is high in dimensionality and has a large amount of redundancy, and direct processing is computationally heavy and difficult to meet real-time requirements; (2) Although deep learning-based target detection models (such as YOLO) have excellent detection speed and accuracy, their decision-making process lacks interpretability, and the output results are difficult to be directly used for high-level decision-making and domain knowledge integration; (3) Existing technical solutions are mostly optimizations of isolated links, such as only performing data dimensionality reduction or only optimizing the detection model, lacking an end-to-end closed-loop analysis process from data preprocessing, real-time perception to pattern cognition.

[0003] Therefore, there is an urgent need in this field for an intelligent visual analysis solution that can effectively compress data, improve processing efficiency, and endow detection results with interpretable pattern discovery capabilities. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the problems of computational redundancy, uninterpretable results and fragmented processes faced by the prior art when processing highly dynamic and multi-target visual data. In order to solve the above problems, an intelligent visual analysis method, device and medium based on dual principal component analysis and YOLO model are provided.

[0005] The object of this invention is achieved in the following manner: A smart visual analysis method based on dual principal component analysis and the YOLO model includes the following steps: S1: Data dimensionality reduction and feature purification: Perform the first principal component analysis on the collected raw visual data to obtain the first feature representation after dimensionality reduction; S2: Real-time target detection and recognition: Input the first feature representation into the YOLO target detection model to perform target detection, and output a structured detection result containing the target location, category and confidence level; S3: Pattern Discovery and Result Interpretation: Based on the structured detection results, a feature matrix is ​​constructed, and a second principal component analysis is performed on the feature matrix to mine potential patterns in the data, and the analysis results including pattern clustering results, anomaly identification results, or quantitative statistical indicators are output.

[0006] In step S1, the raw visual data includes single-frame images, video frame sequences, or spatiotemporal visual feature data; the first principal component analysis includes: The raw visual data is subjected to standardized preprocessing; The number of principal components is determined based on a preset cumulative variance contribution rate threshold in order to reduce the dimensionality of the data; The output is a low-dimensional feature vector or a reconstructed image, which serves as the representation of the first feature.

[0007] The cumulative variance contribution rate threshold is 70% to 95%, preferably 85% to 90%.

[0008] In step S2, the first feature representation is input into the YOLO model in any of the following ways: Method A: Directly input the low-dimensional feature vector into the YOLO model with an adjusted input layer structure; Method B: Input the reconstructed image into the standard YOLO model.

[0009] In method A, adjusting the input layer structure of the YOLO model includes matching the number of input neurons to the dimension of the low-dimensional feature vector.

[0010] In step S3, the specific steps for constructing the feature matrix include: Multiple key feature parameters are extracted from the structured detection results. The key feature parameters include at least two of the following: target bounding box coordinates, motion trajectory, category label, and confidence score. The extracted feature parameters are normalized, and redundant features with variances below a preset threshold are removed. Construct a sample-feature matrix based on the processed feature parameters.

[0011] The extracted key feature parameters are selected according to the application scenario, including: In sports tactical analysis scenarios, this includes player coordinates, movement speed, formation width, and passing network centrality. In intelligent traffic analysis scenarios, data includes traffic density, average vehicle speed, lane occupancy rate, and vehicle spacing variance. In medical image analysis scenarios, this includes the bounding box coordinates, size, and grayscale distribution characteristics of lesions; In industrial visual inspection scenarios, this includes the shape, location, and size characteristics of defects.

[0012] The method is executed in a pipelined parallel manner on a hardware platform that includes a graphics processor, wherein: The first principal component analysis, the YOLO model inference, and the second principal component analysis are distributed to different computational streams of the graphics processor for parallel processing; Asynchronous data transmission between computing tasks is achieved through a data caching pool to reduce end-to-end processing latency; the end-to-end processing latency of the system is controlled to within 50 milliseconds.

[0013] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the intelligent visual analysis method based on dual principal component analysis and the YOLO model.

[0014] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent visual analysis method based on dual principal component analysis and the YOLO model.

[0015] The beneficial effects of this invention are: improved processing efficiency and system robustness: The first PCA significantly reduces the data dimensionality (up to 1 / 3-1 / 5 of the original dimensionality), greatly reducing the computational burden on the subsequent YOLO model and increasing the single-frame processing speed by more than 30%. Simultaneously, the dimensionality reduction process filters out noise such as illumination changes and background jitter, improving target detection accuracy by approximately 7%-12% in complex scenes and enhancing system robustness.

[0016] Enhancing the interpretability and decision support capabilities of results: The second PCA transforms the "black box" detection data output by YOLO into intuitive and interpretable patterns and rules (such as identifying the "triangular attack" tactic and classifying "congested" traffic conditions). In medical image analysis, the consistency between lesion feature clustering results and expert diagnoses can reach over 85%. This makes the analysis results easily understandable to domain experts and directly used for decision-making.

[0017] Achieving a closed-loop intelligent analysis process: This invention constitutes a complete technical closed loop of "data purification → real-time perception → cognitive understanding", which is particularly suitable for application scenarios that require immediate feedback and in-depth review, realizing an upgrade from simple target "detection" to scenario "understanding".

[0018] It possesses high flexibility and applicability: by adapting to different feature extraction strategies and interpretation methods, the method can be seamlessly applied to multiple fields such as sports science, intelligent transportation, medical diagnosis, industrial quality inspection, and security monitoring. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of the intelligent visual analysis method provided in the embodiments of the present invention.

[0020] Figure 2This is a schematic diagram comparing the data distribution before and after the first PCA dimensionality reduction in this embodiment of the invention.

[0021] Figure 3 This is a diagram showing the real-time detection effect of the YOLO model on a basketball game video frame in an embodiment of the present invention.

[0022] Figure 4 This is a clustering visualization of different basketball offensive tactics in the two-dimensional principal component space after the second PCA analysis in this embodiment of the invention.

[0023] Figure 5 This is a bar chart comparing the processing speed and pattern recognition accuracy of the method of this invention with the single YOLO detection method. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0025] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same technical meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0026] A smart visual analysis method based on dual principal component analysis and the YOLO model includes the following steps: S1: Data dimensionality reduction and feature purification: Perform the first principal component analysis on the collected raw visual data to obtain the first feature representation after dimensionality reduction; S2: Real-time target detection and recognition: Input the first feature representation into the YOLO target detection model to perform target detection, and output a structured detection result containing the target location, category and confidence level; S3: Pattern Discovery and Result Interpretation: Based on the structured detection results, a feature matrix is ​​constructed, and a second principal component analysis is performed on the feature matrix to mine potential patterns in the data, and the analysis results including pattern clustering results, anomaly identification results, or quantitative statistical indicators are output.

[0027] In step S1, the raw visual data includes single-frame images, video frame sequences, or spatiotemporal visual feature data; the first principal component analysis includes: The raw visual data is subjected to standardized preprocessing; The number of principal components is determined based on a preset cumulative variance contribution rate threshold in order to reduce the dimensionality of the data; The output is a low-dimensional feature vector or a reconstructed image, which serves as the representation of the first feature.

[0028] The PCA algorithm is used to reduce the dimensionality of the original visual data. The rules for constructing the high-dimensional data matrix are as follows: if the original data is a single-frame image, the matrix is ​​constructed using pixel grayscale values ​​or convolutional features; if it is a video sequence, a temporal matrix is ​​constructed using the spatiotemporal features (coordinates, velocity, acceleration) of the target in each frame. Before PCA processing, Z-score normalization is used to eliminate the influence of dimensions, filtering out noise such as illumination changes and jitter while retaining core information, thus providing stable and clean feature input for subsequent detection.

[0029] The cumulative variance contribution rate threshold is 70% to 95%, preferably 85% to 90%.

[0030] In step S2, the first feature representation is input into the YOLO model in any of the following ways: Method A: Directly input the low-dimensional feature vector into the YOLO model with an adjusted input layer structure; Method B: Input the reconstructed image into the standard YOLO model.

[0031] In method A, adjusting the input layer structure of the YOLO model includes matching the number of input neurons to the dimension of the low-dimensional feature vector.

[0032] In step S3, the specific steps for constructing the feature matrix include: Multiple key feature parameters are extracted from the structured detection results. The key feature parameters include at least two of the following: target bounding box coordinates, motion trajectory, category label, and confidence score. The extracted feature parameters are normalized, and redundant features with variances below a preset threshold are removed. Construct a sample-feature matrix based on the processed feature parameters.

[0033] A secondary PCA analysis was performed on the structured detection results output by YOLO. The feature matrix construction steps were as follows: ① Extract structured features (such as target category proportion, trajectory similarity, and mean confidence score) from the YOLO detection results; ② Normalize the features within the range [0,1]; ③ Remove redundant features with variance less than 0.01 and construct an m×n dimensional feature matrix (m is the number of samples, and n is the number of effective features). This analysis was used to uncover the tactical formations, behavioral patterns, or anomalies contained within the results, which were then output in a visualized and quantifiable form.

[0034] The extracted key feature parameters are selected according to the application scenario, including: In sports tactical analysis scenarios, this includes player coordinates, movement speed, formation width, and passing network centrality. In intelligent traffic analysis scenarios, data includes traffic density, average vehicle speed, lane occupancy rate, and vehicle spacing variance. In medical image analysis scenarios, this includes the bounding box coordinates, size, and grayscale distribution characteristics of lesions; In industrial visual inspection scenarios, this includes the shape, location, and size characteristics of defects.

[0035] The method is executed in a pipelined parallel manner on a hardware platform that includes a graphics processor, wherein: The first principal component analysis, the YOLO model inference, and the second principal component analysis are distributed to different computational streams of the graphics processor for parallel processing; Asynchronous data transmission between computing tasks is achieved through a data caching pool to reduce end-to-end processing latency; the end-to-end processing latency of the system is controlled to within 50 milliseconds.

[0036] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the intelligent visual analysis method based on dual principal component analysis and the YOLO model.

[0037] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent visual analysis method based on dual principal component analysis and the YOLO model.

[0038] Example 1: Tactical Analysis of a Sports Match (Basketball) Data Acquisition: Multiple high-definition cameras were deployed to capture basketball game videos at a frame rate of 25 frames per second and a resolution of 1920×1080.

[0039] First PCA (preprocessing): The video frame sequence is sampled, and the coordinates of all players in each frame are extracted to form a high-dimensional spatiotemporal data matrix; PCA (cumulative variance contribution rate 85%) is performed on this matrix, and the first 3 principal components are selected and interpreted as "formation width", "inside and outside line connection density" and "offensive and defensive transition speed" respectively; the original data is projected onto the space formed by these 3 principal components to obtain the core feature vector after dimensionality reduction.

[0040] YOLO Detection: The PCA-preprocessed features (or the simplified image reconstructed from them) are input into the trained YOLOv8 model (input resolution 640×640, training iterations 10000 times); the model detects and tracks 10 players and the basketball on the court in real time, and outputs the bounding box, identity ID and ball possession status of each target in each frame.

[0041] Second PCA (Pattern Discovery): Collect all attack round data from YOLO throughout the match and construct a feature matrix with features such as "formation principal component score", "player movement speed", and "passing network centrality". Perform a second PCA on this matrix and use the K-means clustering algorithm to analyze the principal component scores. Identify the corresponding tactical patterns based on the feature value range, as shown in the table below.

[0042] Output results: The system successfully identified 5 typical offensive tactical patterns (such as high pick and roll, fast break, triangle offense, etc.), statistically analyzed the frequency of occurrence and scoring efficiency of each pattern, generated a visual tactical report, and displayed heatmaps of different tactics, effect evaluations and targeted suggestions for the coaching team to conduct post-game review and strategy formulation.

[0043] Example 2: Intelligent Traffic Flow Status Analysis Data Acquisition: Video of urban main roads is collected through high-definition road cameras at a frame rate of 30 frames per second and a resolution of 1920×1080 for 24 hours.

[0044] First PCA (preprocessing): Extract features such as traffic density, average vehicle speed, lane occupancy rate, and vehicle spacing variance for each frame to construct a high-dimensional data matrix; perform PCA on the matrix (cumulative variance contribution rate 88%), select the first two principal components, and interpret them as "traffic congestion level" and "driving orderliness" respectively; project the original data onto this two-dimensional principal component space to obtain dimensionality-reduced features.

[0045] YOLO detection: Input the reconstructed simplified image into the YOLOv8 model to detect and track vehicles, pedestrians, and non-motorized vehicles in real time, and output target position, speed, and category information.

[0046] The second PCA (pattern discovery) was performed: a feature matrix was constructed based on "congestion score", "driving orderliness score", "vehicle spacing variance" and "violation frequency". After performing the second PCA, cluster analysis was conducted, and three traffic flow patterns (smooth, slow, and congested) were identified according to the feature value range, as shown in the table below.

[0047] Results Output: Through data integration and time slicing, anomaly detection and labeling, and traffic management suggestion generation rules, a traffic flow status visualization report is generated, which includes heat maps of the distribution patterns in each time period, early warning information of abnormal events (such as congestion origins and illegal gatherings), and traffic management suggestions, providing decision support for traffic management departments. The anomaly detection accuracy rate reaches 92%.

[0048] Experimental Results and Analysis: Test conditions: Hardware platform: NVIDIA RTX 4090 GPU, 128GB RAM; Comparison object: Traditional detection methods using only YOLOv8 (input resolution 640×640, training iterations 10000). Evaluation metrics: single-frame processing speed, target detection accuracy, pattern recognition accuracy, and result interpretability score (1-10 points).

[0049] Experimental data: The method of this invention, through the synergistic effect of dual PCA and YOLO, significantly improves processing speed and result interpretability while maintaining high detection accuracy. It can effectively uncover potential patterns in data and has more comprehensive analytical capabilities and wider application value than traditional methods.

[0050] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several changes and improvements without departing from the overall concept of the present invention, and these should also be considered within the scope of protection of the present invention.

Claims

1. An intelligent visual analysis method based on dual principal component analysis and the YOLO model, characterized in that: Includes the following steps: S1: Data dimensionality reduction and feature purification: Perform the first principal component analysis on the collected raw visual data to obtain the first feature representation after dimensionality reduction; S2: Real-time target detection and recognition: Input the first feature representation into the YOLO target detection model to perform target detection, and output a structured detection result containing the target location, category and confidence level; S3: Pattern Discovery and Result Interpretation: Based on the structured detection results, a feature matrix is ​​constructed, and a second principal component analysis is performed on the feature matrix to mine potential patterns in the data, and the analysis results including pattern clustering results, anomaly identification results, or quantitative statistical indicators are output.

2. The intelligent visual analysis method based on dual principal component analysis and the YOLO model according to claim 1, characterized in that: In step S1, the raw visual data includes single-frame images, video frame sequences, or spatiotemporal visual feature data; The first principal component analysis includes: The raw visual data is subjected to standardized preprocessing; The number of principal components is determined based on a preset cumulative variance contribution rate threshold in order to reduce the dimensionality of the data; The output is a low-dimensional feature vector or a reconstructed image, which serves as the representation of the first feature.

3. The intelligent visual analysis method based on dual principal component analysis and the YOLO model according to claim 2, characterized in that: The cumulative variance contribution rate threshold is 70% to 95%, preferably 85% to 90%.

4. The intelligent visual analysis method based on dual principal component analysis and YOLO model according to claim 1 or 2, characterized in that: In step S2, the first feature representation is input into the YOLO model in any of the following ways: Method A: Directly input the low-dimensional feature vector into the YOLO model with an adjusted input layer structure; Method B: Input the reconstructed image into the standard YOLO model.

5. The intelligent visual analysis method based on dual principal component analysis and the YOLO model according to claim 4, characterized in that: In method A, adjusting the input layer structure of the YOLO model includes matching the number of input neurons to the dimension of the low-dimensional feature vector.

6. The intelligent visual analysis method based on dual principal component analysis and the YOLO model according to claim 1, characterized in that: In step S3, the specific steps for constructing the feature matrix include: Multiple key feature parameters are extracted from the structured detection results. The key feature parameters include at least two of the following: target bounding box coordinates, motion trajectory, category label, and confidence score. The extracted feature parameters are normalized, and redundant features with variances below a preset threshold are removed. Construct a sample-feature matrix based on the processed feature parameters.

7. The intelligent visual analysis method based on dual principal component analysis and the YOLO model according to claim 6, characterized in that: The extracted key feature parameters are selected according to the application scenario, including: In sports tactical analysis scenarios, this includes player coordinates, movement speed, formation width, and passing network centrality. In intelligent traffic analysis scenarios, data includes traffic density, average vehicle speed, lane occupancy rate, and vehicle spacing variance. In medical image analysis scenarios, this includes the bounding box coordinates, size, and grayscale distribution characteristics of lesions; In industrial visual inspection scenarios, this includes the shape, location, and size characteristics of defects.

8. The intelligent visual analysis method based on dual principal component analysis and YOLO model according to claim 1, characterized in that: The method is executed in a pipelined parallel manner on a hardware platform that includes a graphics processor, wherein: The first principal component analysis, the YOLO model inference, and the second principal component analysis are distributed to different computational streams of the graphics processor for parallel processing; Asynchronous data transmission between computing tasks is achieved through a data caching pool to reduce end-to-end processing latency; the end-to-end processing latency of the system is controlled to within 50 milliseconds.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the intelligent visual analysis method based on dual principal component analysis and YOLO model as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the intelligent visual analysis method based on dual principal component analysis and YOLO model as described in any one of claims 1 to 8.

Images