A real-time quality detection method and system based on 5G edge computing

Abstract

The present application relates to a real-time quality detection method and system based on 5G edge computing. The method includes: obtaining dynamic video stream data of real-time video monitoring, and performing low-latency feature extraction of video frames on the dynamic video stream data to obtain video frame feature extraction data, wherein the video frame feature extraction data includes clarity index, color distribution characteristics and brightness fluctuation characteristics; generating a corresponding video frame quality feature matrix according to the clarity index, color distribution characteristics and brightness fluctuation characteristics, and constructing a dynamic quality feature sequence based on timestamp information; performing real-time difference analysis on the quality feature sequences of adjacent video frames through edge nodes to obtain a quality correlation coefficient. The present invention realizes low-latency detection of video quality and abnormality positioning through the collaboration of 5G network and edge computing. The edge node extracts and compresses features such as clarity, color, and brightness of the video frame to reduce bandwidth occupancy and reduce latency.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a real-time quality detection method and system based on 5G edge computing. Background Art

[0002] In today's trend toward digital production and intelligent monitoring, real-time quality inspection methods and systems based on 5G edge computing have emerged as a key technology for ensuring product quality and improving production efficiency. This technology combines the exceptional characteristics of 5G communications with the localized processing advantages of edge computing, enabling real-time, accurate quality inspection across various production processes and video surveillance scenarios.

[0003] Existing quality detection methods often rely on independent evaluation of single-frame pixel features (such as the clarity index), lacking the ability to model temporal variations in video quality. Single-frame analysis cannot capture gradual anomalies such as slow lens deterioration or periodic light source flicker, increasing the rate of missed detections.

[0004] However, the above-mentioned coordinated control scheme determination models provided in the related art all have limitations, resulting in poor regulation effects of the determined coordinated control schemes, and there is room for further improvement. Summary of the Invention

[0005] This application provides a real-time quality detection method and system based on 5G edge computing to at least solve the problems of related technologies.

[0006] In a first aspect, the present application provides a real-time quality detection method based on 5G edge computing, the method comprising:

[0007] Acquire dynamic video stream data of real-time video surveillance, and perform video frame low-latency feature extraction on the dynamic video stream data to obtain video frame feature extraction data, wherein the video frame feature extraction data includes a clarity index, a color distribution feature, and a brightness fluctuation feature;

[0008] Generating a corresponding video frame quality feature matrix according to the clarity index, color distribution feature, and brightness fluctuation feature, and constructing a dynamic quality feature sequence based on timestamp information;

[0009] Performing real-time difference analysis on quality feature sequences of adjacent video frames through edge nodes to obtain quality correlation coefficients, and obtaining influence coefficients of adjacent video frames based on the quality correlation coefficients;

[0010] Constructing a real-time quality association graph based on the influence coefficient, and acquiring video frame data to be evaluated according to the real-time quality association graph, wherein the video frame data to be evaluated includes a frame mark sequence and video frame type information;

[0011] Obtaining a video frame evaluation coefficient according to the frame mark sequence and the video frame type information, and determining whether the video frame evaluation coefficient is less than a preset threshold;

[0012] If it is less than or equal to, the video frame data to be evaluated is determined to be normal;

[0013] If it is greater than, the video frame data to be evaluated is determined to be abnormal.

[0014] Optionally, the step of performing low-latency feature extraction on the dynamic video stream data to obtain video frame feature extraction data includes:

[0015] Obtaining resolution, frame rate, and contrast data of consecutive frames according to a real-time video stream, and calculating a clarity index based on the resolution, frame rate, and contrast data;

[0016] Perform RGB and HSV color space conversion according to the color distribution characteristics of the real-time video stream, and obtain the color temperature deviation value and saturation fluctuation range;

[0017] generating a color distribution feature vector based on the color temperature deviation value and the saturation fluctuation range;

[0018] An average brightness value and a maximum brightness value of a video frame are obtained according to the real-time video stream, and a brightness fluctuation feature is obtained according to the average brightness value and the maximum brightness value.

[0019] Optionally, the step of generating a corresponding video frame quality feature matrix according to the clarity index, color distribution feature, and brightness fluctuation feature, and constructing a dynamic quality feature sequence based on timestamp information includes:

[0020] Assign a unique timestamp to each frame of video and sort the quality feature matrix in the order of timestamps;

[0021] Obtain the time interval between adjacent frames, and perform sliding window processing on the quality feature matrix based on the time interval to generate a local quality feature subsequence;

[0022] Calculate the mean, variance and change rate of each local quality feature subsequence to obtain the dynamic quality feature vector;

[0023] Combine the dynamic quality feature vector with the original quality feature matrix to generate a dynamic quality feature sequence;

[0024] Perform discrete Fourier transform on the dynamic quality feature sequence to extract periodic quality fluctuation characteristics;

[0025] The periodic quality fluctuation characteristics are integrated with the dynamic quality feature sequence to form a complete dynamic quality feature sequence.

[0026] Optionally, the step of performing real-time difference analysis on quality feature sequences of adjacent video frames through edge nodes to obtain quality correlation coefficients, and obtaining influence coefficients of adjacent video frames based on the quality correlation coefficients includes:

[0027] In the edge node, the pre-trained difference analysis of adjacent video frames is performed to obtain the initial model parameters;

[0028] Perform frame-by-frame difference calculation on the quality feature sequences of adjacent video frames to generate a feature difference matrix;

[0029] Obtain the relative change rate of each feature item based on the feature difference matrix, and obtain the probability distribution of each feature item;

[0030] Obtaining a characteristic anomaly entropy value according to the probability distribution;

[0031] Generate a quality correlation coefficient matrix according to the initialization model parameters and characteristic anomaly entropy values;

[0032] The quality correlation coefficient matrix is ​​normalized to obtain the influence coefficients of adjacent video frames.

[0033] Optionally, the step of constructing a real-time quality association graph based on the influence coefficient includes:

[0034] Each video frame is represented as a node, and the influence coefficients of adjacent video frames are represented as edge weights;

[0035] Construct an initial association graph based on edge weights and mark the abnormal video frame sequence position of each edge;

[0036] Obtaining the associated path of each abnormal video frame according to the position of the abnormal video frame sequence;

[0037] Obtain the sum of the delays for each associated path, and then find the path with the minimum propagation delay, which will be used as the critical path.

[0038] A real-time quality association diagram is generated through topological structure analysis according to the critical path.

[0039] Optionally, the step of acquiring the video frame data to be evaluated according to the real-time quality association graph includes:

[0040] Obtaining video frames of a critical path according to the real-time quality association graph, and recording them as candidate frames to be evaluated;

[0041] Parse the original annotation data of the candidate frame to be evaluated to obtain the frame label sequence;

[0042] Acquiring a shooting scene, device type, and content features of the video frame, and generating video frame type information according to the shooting scene, device type, and content features of the video frame;

[0043] Perform local sliding window processing on the dynamic quality feature sequence of the candidate frame to be evaluated to extract the time correlation features;

[0044] Combine the confidence values ​​of the frame label sequence to generate a label confidence feature vector;

[0045] The time correlation feature is fused with the label confidence feature vector to construct a dynamic feature correlation matrix;

[0046] The video frame data to be evaluated is obtained according to the video frame type information and the dynamic feature association matrix.

[0047] Optionally, the step of obtaining a video frame evaluation coefficient according to the frame marker sequence and the video frame type information includes:

[0048] Marking the frame sequence to obtain marking information, wherein the marking information includes normal marking, suspicious marking and abnormal marking;

[0049] Classify and encode the video frame type information to generate a type feature vector;

[0050] Construct an evaluation input matrix based on the label information and type feature vector;

[0051] generating preliminary evaluation coefficients according to the evaluation input matrix;

[0052] A final video frame evaluation coefficient is obtained according to the preliminary evaluation coefficient.

[0053] In a second aspect, the present application provides a real-time quality detection system based on 5G edge computing, the system comprising:

[0054] A video processing module is used to obtain dynamic video stream data of real-time video surveillance and perform low-latency feature extraction of video frames on the dynamic video stream data to obtain video frame feature extraction data, wherein the video frame feature extraction data includes a clarity index, color distribution characteristics, and brightness fluctuation characteristics;

[0055] A sequence construction module is used to generate a corresponding video frame quality feature matrix according to the clarity index, color distribution characteristics and brightness fluctuation characteristics, and to construct a dynamic quality feature sequence based on timestamp information;

[0056] A difference analysis module is used to perform real-time difference analysis on quality feature sequences of adjacent video frames through edge nodes to obtain quality correlation coefficients, and obtain influence coefficients of adjacent video frames based on the quality correlation coefficients;

[0057] an association graph construction module, configured to construct a real-time quality association graph based on the influence coefficient, and obtain video frame data to be evaluated according to the real-time quality association graph, wherein the video frame data to be evaluated includes a frame mark sequence and video frame type information;

[0058] An evaluation decision module, configured to obtain a video frame evaluation coefficient based on the frame mark sequence and the video frame type information, and determine whether the video frame evaluation coefficient is less than a preset threshold;

[0059] If it is less than or equal to, the video frame data to be evaluated is determined to be normal;

[0060] If it is greater than, the video frame data to be evaluated is determined to be abnormal.

[0061] Preferably, the sequence building block comprises:

[0062] a clarity extraction unit, configured to obtain resolution, frame rate, and contrast data of consecutive frames according to a real-time video stream, and calculate a clarity index based on the resolution, frame rate, and contrast data;

[0063] A color analysis unit, configured to perform RGB and HSV color space conversion according to the color distribution characteristics of the real-time video stream, and obtain a color temperature deviation value and a saturation fluctuation range;

[0064] A feature generation unit, configured to generate a color distribution feature vector based on the color temperature deviation value and the saturation fluctuation range;

[0065] The brightness analysis unit is used to obtain an average brightness value and a maximum brightness value of a video frame according to the real-time video stream, and obtain a brightness fluctuation feature according to the average brightness value and the maximum brightness value.

[0066] In a third aspect, the present application provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the real-time quality detection method based on 5G edge computing provided in the first aspect above is implemented.

[0067] Compared with the related art, the present application provides at least the following technical effects:

[0068] By synergizing 5G networks with edge computing, low-latency video quality detection and anomaly location are achieved. Edge nodes extract and compress video frame features such as clarity, color, and brightness, reducing bandwidth usage and latency. Multi-dimensional integration, including dynamic quality feature sequences and real-time quality correlation graphs, improves anomaly detection accuracy. Using graph theory to construct correlation graphs, combined with critical path analysis, shortens the time required to locate anomaly sources. Video frame type information enables scene-adaptive assessment, dynamically adjusting thresholds to adapt to different environments.

[0069] The details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0071] Figure 1 Schematic diagram of a method flow according to an embodiment of the present invention.

[0072] Figure 2 FIG. 1 is a schematic diagram of a system structure according to an embodiment of the present invention.

[0073] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of the present application. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is described and illustrated below in conjunction with the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely used to explain this application and are not intended to limit this application. Based on the embodiments provided in this application, all other embodiments obtained by those of ordinary skill in the art without making any creative efforts are within the scope of protection of this application.

[0075] Obviously, the drawings described below are merely examples or embodiments of the present application. Those skilled in the art can, without inventive effort, apply the present application to other similar scenarios based on these drawings. Furthermore, it is also understood that, although the effort involved in such a development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, changes in design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as an insufficiency of the content disclosed in this application.

[0076] References to "embodiments" in this application mean that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it refer to independent or alternative embodiments that are mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described in this application may be combined with other embodiments unless there is a conflict.

[0077] Unless otherwise defined, technical or scientific terms used herein shall have the ordinary meaning as understood by persons of ordinary skill in the art to which this application belongs. The terms "a," "an," "an," "the," and similar expressions used herein do not denote quantitative limitations and may refer to either the singular or the plural. The terms "comprise," "include," "have," and any variations thereof, used herein, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or modules (units) is not limited to the listed steps or units but may also include steps or units not listed, or may include other steps or units inherent to the process, method, product, or apparatus. The terms "connected," "connected," "coupled," and similar expressions used herein are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. As used herein, "plurality" means two or more. "And / or" describes an association between associated objects, indicating that three possible relationships exist. For example, "A and / or B" may mean: A exists alone; A and B exist simultaneously; or B exists alone. The character " / " generally indicates that the objects before and after are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific order for the objects.

[0078] Example 1

[0079] An embodiment of the present invention provides a real-time quality detection method based on 5G edge computing. Figure 1 This is a flow chart according to an exemplary embodiment. It includes:

[0080] S1. Acquire dynamic video stream data of real-time video surveillance, and perform video frame low-latency feature extraction on the dynamic video stream data to obtain video frame feature extraction data, wherein the video frame feature extraction data includes a clarity index, a color distribution feature, and a brightness fluctuation feature;

[0081] S2. Generate a corresponding video frame quality feature matrix according to the clarity index, color distribution characteristics, and brightness fluctuation characteristics, and construct a dynamic quality feature sequence based on timestamp information;

[0082] S3. Performing real-time difference analysis on quality feature sequences of adjacent video frames through edge nodes to obtain quality correlation coefficients, and obtaining influence coefficients of adjacent video frames based on the quality correlation coefficients;

[0083] S4. Constructing a real-time quality association graph based on the influence coefficient, and obtaining video frame data to be evaluated according to the real-time quality association graph, wherein the video frame data to be evaluated includes a frame marker sequence and video frame type information;

[0084] S5. Obtaining a video frame evaluation coefficient according to the frame mark sequence and the video frame type information, and determining whether the video frame evaluation coefficient is less than a preset threshold;

[0085] If it is less than or equal to, the video frame data to be evaluated is determined to be normal;

[0086] If it is greater than, the video frame data to be evaluated is determined to be abnormal.

[0087] As described in S1-S5 above, the present invention acquires dynamic video stream data in real time via a 5G network and employs a low-latency algorithm to extract features from video frames. Edge nodes rapidly analyze video frame parameters such as resolution, frame rate, and contrast using lightweight models (such as convolutional neural networks), and calculate the sharpness index using frequency domain analysis. Simultaneously, color distribution features (such as color temperature deviation and saturation fluctuation) are extracted through RGB and HSV color space conversion, and brightness fluctuation features are calculated using the mean and maximum brightness values. This method reduces feature dimensionality through normalization and principal component analysis (PCA) to generate a standardized quality feature matrix. Compared to traditional centralized cloud-based processing, localized feature extraction at the edge node significantly reduces transmission latency and ensures real-time video quality assessment. For example, in industrial monitoring scenarios, if a camera lens is blurred due to lens contamination, the sudden drop in sharpness index can be captured in real time, preventing misidentification during subsequent defect detection.

[0088] Then, a dynamic quality feature sequence is constructed using timestamp information to capture the temporal variation trends of the video stream. Specifically, the edge node assigns a millisecond-level timestamp to each frame and performs a local analysis of the feature matrices of adjacent frames using a sliding window technique. The mean, variance, and rate of change are calculated to generate a dynamic feature vector. Periodic fluctuation features are further extracted using a discrete Fourier transform (DFT), and static and dynamic features are ultimately fused to form a complete dynamic quality feature sequence. This solves the problem that traditional static feature analysis cannot reflect quality variation trends and is particularly suitable for detecting brightness fluctuations caused by changes in ambient light during nighttime monitoring. For example, when the light source on a factory production line periodically flickers due to voltage instability, the dynamic feature sequence can accurately capture the periodic pattern of brightness fluctuations, providing a reliable basis for subsequent anomaly determination.

[0089] Next, edge nodes perform real-time differential analysis on the dynamic quality feature sequences of adjacent frames to quantify inter-frame quality degradation relationships. Specifically, a frame-by-frame differential operation is used to generate a feature difference matrix. Feature anomaly entropy is calculated using a probability distribution, and the time interval is combined to generate a quality correlation coefficient. Network bandwidth utilization and transmission delay are further incorporated, and influence coefficients are generated through weighted fusion. This method reduces cloud-side reliance by localizing differential analysis. For example, in a 5G edge server, single-frame processing time can be reduced to less than 5ms, significantly improving real-time performance.

[0090] Next, a real-time quality association graph is constructed based on the impact coefficient, and high-risk video frames are screened. Specifically, each frame is represented as a node, and the impact coefficient is used as the edge weight. A breadth-first search is used to identify anomaly propagation paths. Metadata for the frames to be evaluated is generated by combining the label sequence (normal / suspicious / abnormal) and video frame type information (such as industrial monitoring frames and live broadcast frames). High-risk frames are screened using a dynamic feature association matrix, and anomaly confidence scores are calculated. This method uses graph theory optimization to accurately identify anomaly propagation chains.

[0091] Finally, by integrating the tag sequence, type information, and dynamic feature vectors, a random forest model is used to calculate the video frame evaluation coefficient. This coefficient is then compared with a preset threshold to identify anomalies. Specifically, a dynamic threshold mechanism calculates the rolling mean and standard deviation based on historical data, adaptively adjusting the threshold range to avoid false positives. For example, the threshold is lower during stable daylight hours and automatically increases during nighttime periods of high light fluctuations.

[0092] In one embodiment, the step of performing low-latency feature extraction on the dynamic video stream data to obtain video frame feature extraction data includes:

[0093] S101, obtaining resolution, frame rate, and contrast data of consecutive frames according to a real-time video stream, and calculating a clarity index based on the resolution, frame rate, and contrast data;

[0094] S102, performing RGB and HSV color space conversion according to the color distribution characteristics of the real-time video stream, and obtaining a color temperature deviation value and a saturation fluctuation range;

[0095] S103, generating a color distribution feature vector based on the color temperature deviation value and the saturation fluctuation range;

[0096] S104 , obtaining an average brightness value and a maximum brightness value of a video frame according to the real-time video stream, and obtaining a brightness fluctuation feature according to the average brightness value and the maximum brightness value.

[0097] In summary, the present invention achieves quantitative analysis of video frame clarity by extracting the resolution, frame rate, and contrast data of the video stream. Basic parameters such as resolution and frame rate are obtained from the video stream metadata. The Canny edge detection operator is used to process the video frame grayscale matrix. The edge strength is calculated by calculating the square root of the sum of the squares of the x- and y-direction gradients. The edge pixel density per unit area is then calculated to generate a clarity index. Characterizing image clarity by locally calculating edge density on the edge side can effectively identify blur anomalies such as lens contamination and inaccurate focus.

[0098] To accurately identify color anomalies in the video, the RGB color space is converted to HSV space, separating the brightness and color components to calculate color temperature deviation and saturation fluctuations. The hue, saturation, and brightness components are obtained using the HSV conversion formula. The hue value is converted to color temperature and compared with the standard color temperature to obtain the deviation value. The standard deviation of the saturation values ​​across multiple consecutive frames is calculated to characterize the fluctuation range. This solves the latency issue inherent in traditional color analysis methods, which rely on cloud-based processing. Real-time conversion and statistics at the edge can capture color shifts caused by light source changes or color mutations due to filter failures.

[0099] Then, to achieve structured expression of color features, the color temperature deviation and saturation fluctuation are normalized to generate a feature vector. The dimensional differences between different features are eliminated through the Z-score normalization method, and the standardized parameters are combined into a two-dimensional feature vector as a quantitative representation of color anomalies. This operation solves the analysis difficulties caused by the unstructured nature of traditional features. The vector form facilitates subsequent machine learning model training and matching with historical anomaly samples. The feature vector dimension is compressed to two dimensions, reducing the amount of data while retaining key color information. It supports rapid matching of anomaly patterns through Euclidean distance, provides a standardized data foundation for cross-frame color anomaly correlation analysis, and improves the efficiency of edge-side feature storage and transmission.

[0100] Finally, brightness anomaly detection is achieved by calculating the average and maximum brightness of video frames and combining them with sliding window analysis. The average and maximum brightness values ​​are calculated from the pixel data, and the brightness change rate is calculated using a 20-frame sliding window to capture brightness anomalies such as lens occlusion and sudden changes in light sources. This method overcomes the problem of traditional single-point brightness detection's inability to identify temporal changes. It improves the ability to detect gradual or sudden brightness anomalies through temporal analysis. The brightness detection range covers the full dynamic range, capable of identifying even subtle brightness changes. The response time to scenarios such as sudden strong light interference meets real-time monitoring requirements, and the false alarm rate is within an acceptable industry standard.

[0101] In one embodiment, the step of generating a corresponding video frame quality feature matrix according to the clarity index, color distribution feature, and brightness fluctuation feature, and constructing a dynamic quality feature sequence based on timestamp information includes:

[0102] S201, assigning a unique timestamp to each frame of video, and sorting the quality feature matrix according to the order of the timestamps;

[0103] S202, obtaining the time interval between adjacent frames, and performing sliding window processing on the quality feature matrix based on the time interval to generate a local quality feature subsequence;

[0104] S203, calculating the mean, variance, and change rate of each local quality feature subsequence to obtain a dynamic quality feature vector;

[0105] S204, combining the dynamic quality feature vector with the original quality feature matrix to generate a dynamic quality feature sequence;

[0106] Perform discrete Fourier transform on the dynamic quality feature sequence to extract periodic quality fluctuation characteristics;

[0107] S205. Merge the periodic quality fluctuation feature with the dynamic quality feature sequence to form a complete dynamic quality feature sequence.

[0108] In summary, this invention establishes a temporal correlation basis for video frames by assigning a unique timestamp to each frame and sequentially arranging the quality feature matrix. Timestamps are extracted from the video stream, generating a time identifier for each frame. The single-frame quality feature matrix (including clarity, color, and brightness features) is then arranged in ascending order by timestamp, forming an ordered temporal matrix. This overcomes the lack of a temporal dimension in traditional single-frame analysis, enabling subsequent analysis to capture temporal variations in video quality. Timestamp synchronization errors are kept within industry standards, ensuring temporal consistency across multiple frames and providing an ordered data structure for sliding window processing and cross-frame difference analysis.

[0109] Then, a sliding window process is performed on the quality feature matrix based on the time interval between adjacent frames. The window parameters are determined by calculating the frame interval, and a fixed window size and overlapping step size are used to generate local quality feature subsequences. For example, the feature matrix is ​​truncated with an appropriate window size and step size to form a set of continuous overlapping subsequences. This method simulates the short-term memory properties of human vision and can effectively identify short-term anomalies such as sudden brightness changes caused by sudden occlusion.

[0110] Next, statistics are calculated for each local quality feature subsequence to generate a dynamic feature vector representing short-term quality changes. Specifically, the mean, variance, and rate of change are calculated for each subsequence, reflecting the short-term quality level, feature stability, and speed of quality fluctuation, respectively. These three are combined to form a dynamic feature vector. This vector expands single-frame features into short-term dynamic features, effectively quantifying the quality degradation process. For example, the rate of change can identify a slow decline in clarity, while variance analysis can determine the severity of brightness fluctuations. The calculation of statistics combines feature aggregation over the time dimension, resolving the problem that traditional single-frame analysis cannot reflect quality change trends, enabling edge nodes to perform dynamic modeling of short-term features locally.

[0111] Next, the dynamic feature vectors are combined with the original quality feature matrix to generate a spatiotemporal feature sequence, which is then subjected to a discrete Fourier transform. This Fourier transform converts the time-domain features into a frequency-domain representation, extracting periodic fluctuation characteristics, such as those associated with periodic vibration blurring caused by the operation of industrial equipment. The spatiotemporal feature sequence preserves both single-frame details and short-term fluctuations, while frequency-domain analysis further analyzes long-term patterns. This combination addresses the inability of traditional methods to identify regular interference sources.

[0112] Finally, a complete dynamic quality feature sequence is formed by screening effective frequency domain features and fusing them with the dynamic feature sequence. First, a frequency threshold is set to extract meaningful periodic features. The frequency domain features are then converted to a time domain representation and fused with the original sequence, ensuring that the final sequence contains both short-term fluctuations and long-term patterns. This fusion operation addresses the limitations of single-domain analysis, such as single-domain analysis. For example, it can simultaneously reflect different anomaly patterns, such as "current frame blur" and "once-per-hour blur." The integration of periodic features significantly reduces the system's underreporting rate for regular anomalies. The complete dynamic feature sequence provides multi-dimensional feature support for the subsequent construction of a real-time quality correlation map.

[0113] In one embodiment, the step of performing real-time difference analysis on quality feature sequences of adjacent video frames through edge nodes to obtain quality correlation coefficients, and obtaining influence coefficients of adjacent video frames based on the quality correlation coefficients includes:

[0114] S301, performing pre-trained difference analysis on adjacent video frames in the edge node to obtain initialization model parameters;

[0115] S302, performing frame-by-frame difference operation on quality feature sequences of adjacent video frames to generate a feature difference matrix;

[0116] S303, obtaining the relative change rate of each feature item based on the feature difference matrix, and obtaining the probability distribution of each feature item;

[0117] S304, obtaining a characteristic abnormality entropy value according to the probability distribution;

[0118] S305, generating a quality correlation coefficient matrix according to the initialization model parameters and characteristic anomaly entropy values;

[0119] S306 : Normalize the quality correlation coefficient matrix to obtain influence coefficients of adjacent video frames.

[0120] In summary, this invention reduces cloud dependency and improves real-time performance by locally deploying lightweight models on edge nodes. In a specific implementation, edge nodes load a pretrained convolutional neural network (CNN) consisting of three 3×3 convolutional layers, with the number of channels increasing from 64 to 128 to 256, and an output feature dimension of 256. Model parameters are initialized using the Xavier method to avoid vanishing or exploding gradients. These parameters are then pre-delivered to edge nodes via the 5G network and stored in NVRAM to ensure fast loading. This method reduces computational overhead through a lightweight CNN, ensuring that edge nodes complete feature extraction within a preset timeframe.

[0121] Next, a frame-by-frame differential operation is performed on the quality feature sequences of adjacent frames to generate a feature difference matrix. After obtaining the quality feature sequences of two adjacent frames, the Euclidean distance is calculated for the feature vectors at each time point to construct a matrix reflecting the degree of feature difference between the frames. The dimensions of this matrix correspond to the time series length and feature dimension, and its element values ​​represent the strength of feature differences between adjacent frames at corresponding time points. In this way, abstract quality feature changes are converted into structured matrix data, facilitating subsequent analysis of temporal patterns and abnormal fluctuations in feature changes.

[0122] Next, the relative rate of change of each feature item is calculated and its probability distribution is statistically analyzed to quantify the degree and regularity of feature variation. For each feature dimension of the difference matrix, the relative rate of change between adjacent time points is calculated to reflect the temporal fluctuations of feature differences. By counting the frequency of occurrence of each change rate, a probability distribution is generated to characterize the uncertainty and regularity of feature variation. This process transforms the numerical difference matrix into probabilistic statistical features, effectively describing the randomness and trend of feature variation, and addressing the problem of traditional methods that focus solely on absolute differences while ignoring the regularity of variation.

[0123] The feature anomaly entropy is calculated based on the probability distribution of the feature item's rate of change to quantify the uncertainty of the feature change. Entropy calculation is based on the probability of occurrence of each rate of change in the probability distribution and is derived using the information entropy formula. A larger entropy value indicates greater uncertainty in the feature change, and therefore a greater likelihood of an anomaly. This entropy, as a quantitative indicator of the degree of feature anomaly, effectively integrates the magnitude of feature change with the characteristics of the probability distribution, overcoming the limitations of traditional single threshold judgments. Through entropy analysis, edge nodes can assess the degree of anomaly in feature changes from an information-theoretic perspective.

[0124] Then, combining the pre-trained model parameters with the feature anomaly entropy values, this step generates a quality correlation coefficient matrix, representing the strength of the quality correlation between adjacent video frames. By weightedly integrating the feature difference representations extracted by the pre-trained model with the anomaly uncertainty reflected by the entropy values, a multidimensional correlation coefficient matrix is ​​constructed. The matrix element values ​​comprehensively consider the absolute strength, variation patterns, and anomaly probability of feature differences, resulting in a comprehensive quantification of inter-frame quality correlations. This matrix not only reflects the direct feature differences between adjacent frames but also incorporates the temporal correlation patterns learned by the model, addressing the single-dimensionality of correlation analysis in traditional methods. The quality correlation coefficient matrix provides the basis for the subsequent calculation of impact coefficients, enabling edge nodes to assess inter-frame quality impacts from multiple dimensions and providing more precise parameter support for anomaly propagation analysis.

[0125] Finally, the quality correlation coefficient matrix is ​​normalized to generate standardized influence coefficients for adjacent video frames, facilitating cross-scenario comparison and application. By mapping the element values ​​of the correlation coefficient matrix to a reasonable range, dimensional differences are eliminated, and a unified quantitative index for the influence coefficient is formed. Normalization makes the influence coefficients comparable and can be directly used to construct edge weights for real-time quality correlation graphs, supporting graph-theory-based analysis of abnormal propagation paths. This step resolves the analytical difficulties caused by the different dimensions of traditional correlation parameters, making the influence coefficients generated by edge nodes widely applicable to different monitoring scenarios. The normalized influence coefficients accurately reflect the degree of quality impact of adjacent frames, providing a standardized quantitative basis for subsequent abnormality source location and propagation path identification.

[0126] In one embodiment, the step of constructing a real-time quality association graph based on the influence coefficient includes:

[0127] S401, representing each video frame as a node, and representing the influence coefficients of adjacent video frames as edge weights;

[0128] S402: construct an initial association graph based on edge weights, and mark the abnormal video frame sequence position of each edge;

[0129] S403, obtaining an associated path of each abnormal video frame according to the position of the abnormal video frame sequence;

[0130] S404: Obtain the total delay of each associated path, obtain the path with the minimum propagation delay, and use it as the key path;

[0131] S405 : Generate a real-time quality association graph based on the critical path through topological structure analysis.

[0132] In summary, the present invention constructs a basic graph structure associated with video quality by abstracting video frames into graph nodes and using influence coefficients as edge weights. Specifically, each video frame is mapped to a graph node, and the node attributes include timestamps and quality features. The weights of directed edges between adjacent frames are set to the influence coefficients output in step S306, forming a weighted directed graph model. The adjacency matrix is ​​used to store the graph structure to ensure efficient node query and edge weight update. This operation converts the temporal association of video frames into node and edge relationships in graph theory, solving the problem of lack of visualization and quantitative representation in traditional temporal analysis, so that the degree of quality impact between frames can be intuitively reflected through edge weights;

[0133] Next, an initial association graph is constructed based on edge weights, annotating historical anomaly locations. Nodes and edges are added sequentially according to the video frame time sequence to form a chained initial graph. The historical anomaly records are then traversed, the timestamp locations of the edges that caused the anomalies are marked, and the anomaly probability of each edge is calculated as an attribute. The initial graph reflects the inter-frame associations under normal conditions, and the historical anomaly annotations provide prior knowledge for subsequent path search, ensuring that the graph structure simultaneously captures current quality associations and historical anomaly patterns.

[0134] Then, possible anomaly-associated paths are obtained through graph traversal. First, the node position of the anomaly frame in the graph is located. A breadth-first search is used to traverse backward from this node, limiting the search depth and filtering low-weight edges. Only paths with an impact coefficient greater than a threshold are retained. The reverse search strategy conforms to the anomaly tracing logic, inferring the possible source from the current anomaly. Depth and weight filtering balance search efficiency and effectiveness, avoiding interference from invalid paths. This solves the inefficiency of traditional manual analysis of anomaly propagation paths. The algorithm automatically generates a set of possible propagation paths, providing candidates for subsequent key path determination, making the acquisition of anomaly-associated paths automated and real-time.

[0135] Next, the critical path is determined by calculating propagation delay. For each valid path, the transmission delay of each edge is accumulated. Using the Dijkstra algorithm, the path with the shortest delay is selected as the critical path. If equal-delay paths exist, they are sorted by edge anomaly probability. Propagation delay reflects the actual propagation speed of the anomaly. The path with the shortest delay is most likely the true propagation chain. Anomaly probability serves as an auxiliary indicator to improve reliability. This addresses the ambiguity in path priority determination in traditional methods. By using quantitative indicators to clearly identify the critical path, the source of the anomaly is more accurately located, providing a reliable path basis for subsequent root cause analysis and improving the efficiency and accuracy of anomaly handling.

[0136] Finally, this step generates a real-time quality association graph through topological analysis. The topological features of the critical path are extracted, and the graph structure is maintained in combination with a dynamic update strategy. New frames are added in sequence, and the critical path is recalculated and visually annotated when an anomaly is updated. The generated graph structure contains node quality status and critical path annotations, and the quality anomalies and the degree of impact are intuitively displayed through node color and edge width. Topological analysis refines the structural features of the critical path, and dynamic updates ensure that the graph reflects the quality association of the video stream in real time. The visual output facilitates operation and maintenance personnel to quickly locate anomalies. This step solves the problem of the lack of intuitive display of traditional association analysis. The generated real-time quality association graph provides visual support for locating the root cause of video quality, upgrading the anomaly propagation analysis from data-driven to graph visualization-driven, greatly improving the efficiency of anomaly handling.

[0137] In one embodiment, the step of obtaining the video frame data to be evaluated according to the real-time quality correlation graph includes:

[0138] S406, obtaining video frames of a critical path according to the real-time quality association graph, and recording them as candidate frames to be evaluated;

[0139] S407, parsing the original annotation data of the candidate frame to be evaluated to obtain a frame label sequence;

[0140] S408: Acquire the shooting scene, device type, and content features of the video frame, and generate video frame type information according to the shooting scene, device type, and content features of the video frame;

[0141] S409, performing local sliding window processing on the dynamic quality feature sequence of the candidate frame to be evaluated to extract time correlation features;

[0142] S410, generating a marker confidence feature vector based on the confidence values ​​of the frame marker sequence;

[0143] S411, fusing the time correlation feature with the tag confidence feature vector to construct a dynamic feature association matrix;

[0144] S412: Obtain the video frame data to be evaluated according to the video frame type information and the dynamic feature association matrix.

[0145] In summary, the present invention extracts video frames on key propagation paths from the real-time quality association graph as candidate frames to be evaluated, in order to focus on high-risk abnormal propagation links. Specifically, by locating the key path (such as the shortest delay path) in the association graph, all video frame nodes on the path are obtained, and the candidate frames are arranged in descending order according to the influence coefficient weights of the nodes, with video frames with high influence coefficients being included first. This operation is based on the temporal correlation of abnormal propagation, reducing the number of frames to be evaluated to a small proportion of the original video stream, solving the problem of low efficiency of traditional full-scale evaluation, allowing high-risk frames to be processed first, and ensuring the efficiency of abnormal tracing while significantly reducing the amount of calculation.

[0146] Next, the original annotated data of the candidate frame to be evaluated is parsed. Manual or historical automatic annotations are read from the edge node cache, and the tag sequence (e.g., normal / blurred) and corresponding confidence scores (if available) are extracted. By parsing the annotated data, the historical quality status of the candidate frame is determined. The confidence level is used to quantify the reliability of the annotations, preventing misjudgments due to incorrect annotations. This process provides a reference for subsequent evaluation, ensuring the validity of the annotation information and improving the accuracy of the evaluation results.

[0147] Next, multi-dimensional metadata is integrated to generate video frame type information to adapt to the evaluation requirements of different scenarios. The shooting scene (e.g., industrial production line, urban road) and device type are extracted from the video stream. A lightweight CNN is used to extract content features (e.g., moving objects, texture complexity), and these scene, device, and content features are converted into encoding vectors. The generation of type information enables dynamic adjustment of evaluation criteria based on the scenario (e.g., industrial scenarios require higher clarity), addressing the problem of traditional evaluation methods ignoring scene differences and improving the adaptability of evaluation in different environments.

[0148] Next, a sliding window process is performed on the dynamic quality feature sequence of the candidate frames to be evaluated to extract temporal correlation features. A sliding window of appropriate size is used to traverse the feature sequence, and the Pearson correlation coefficient of the features within the window is calculated to generate a correlation matrix. Principal component analysis is then used to reduce the dimensionality and obtain the feature vector. This process effectively identifies gradual anomalies, such as a gradual decrease in clarity, by analyzing the degree of feature correlation between adjacent frames. Compared to traditional single-frame analysis, it significantly improves the ability to capture temporal anomalies.

[0149] The tag information is then combined with the confidence quantification to generate a feature vector representing the tag reliability. The tag confidence values ​​are standardized, and the tag status (normal / suspicious / abnormal) is encoded as a vector. The two are then weighted and fused to form a comprehensive feature vector. Through standardization and encoding fusion, the reliability of the tag is quantified, avoiding misleading low-confidence tags. This addresses the tag uncertainty issue in traditional assessments, provides a more comprehensive representation of tag information, and provides reliable tag dimension data for subsequent feature fusion.

[0150] Next, the temporal correlation features are fused with the labeling confidence vector to construct a dynamic feature association matrix. After aligning the two feature dimensions, the matrix is ​​arranged chronologically, and weights are assigned to different feature columns to highlight key dimensions. This matrix structure characterizes the temporal correlation and labeling reliability of candidate frames, enabling efficient cross-frame anomaly pattern recognition. Compared to single feature analysis, this significantly improves anomaly pattern recognition efficiency and provides structured feature association data for subsequent comprehensive evaluation.

[0151] Finally, the video frame type information is fused with the dynamic feature association matrix to generate standardized data for evaluation. The type encoding vector and feature matrix are concatenated, standardized, and then packaged into a unified format. This fusion of multi-dimensional features ensures that the evaluation data includes information such as scene, device, time sequence, and tagging, resolving the issue of traditional evaluations with a single feature dimension. Standardized packaging ensures efficient data processing, ultimately improving the accuracy of anomaly assessments.

[0152] In one embodiment, the step of obtaining a video frame evaluation coefficient according to the frame marker sequence and the video frame type information includes:

[0153] S501, obtaining marking information from a frame marking sequence, wherein the marking information includes a normal mark, a suspicious mark, and an abnormal mark;

[0154] S502, classify and encode the video frame type information to generate a type feature vector;

[0155] S503, constructing an evaluation input matrix based on the tag information and the type feature vector;

[0156] S504, generating preliminary evaluation coefficients according to the evaluation input matrix;

[0157] S505 : Obtain final video frame evaluation coefficients according to the preliminary evaluation coefficients.

[0158] In summary, the present invention extracts labeling information from the raw annotated data of the video frame to be evaluated, including three labeling states: normal, suspicious, and abnormal, and their corresponding confidence scores. By parsing the annotated data obtained in step S407, the labeling state is converted into a three-dimensional vector (e.g., a normal label corresponds to [1, 0, 0]), and a confidence value is extracted to quantify the reliability of the labeling. This operation converts historical quality assessment results into structured features, providing a temporal reference for the current frame evaluation and avoiding misjudgments caused by low-confidence labels. Through labeling state encoding and confidence filtering, historical labeling information is effectively integrated, improving the reliability of the evaluation basis.

[0159] Next, the video frames' shooting scenes, device types, and content features are classified and encoded to generate a type feature vector. The scene (e.g., an industrial production line), device model (e.g., a Hikvision model), and content features (e.g., the proportion of moving objects) obtained from step S408 are converted into a feature vector of uniform dimension through one-hot encoding and normalization. The type feature encoding defines the quality assessment criteria for different scenarios. For example, industrial scenes require higher clarity, and the device type influences the brightness threshold setting. This process enables the assessment model to adapt to the needs of different scenarios, reducing misjudgments due to scene variations.

[0160] The tag information is then fused with the type feature vector to construct a multidimensional evaluation input matrix. Through dimensional alignment, the 3D tag state vector and the N-dimensional type feature vector are concatenated into a matrix of uniform dimensions. Initial weights are assigned to each feature column based on historical data (e.g., anomaly tags are weighted higher than common type features). The evaluation matrix integrates tag timing and scene attributes to form a multidimensional evaluation foundation. The weight assignments reflect the differential contributions of different features to quality assessment. This matrix structure transcends the limitations of traditional single-feature evaluation, encompassing more evaluation dimensions. This provides comprehensive feature input for subsequent intelligent evaluation, improving the accuracy and comprehensiveness of the evaluation results.

[0161] Next, a lightweight neural network processes the evaluation input matrix to generate preliminary evaluation coefficients. A two-layer fully connected neural network (input layer = N+3, hidden layer = 8, output layer = 1) is used to convert the matrix features into preliminary coefficients in the range [0, 1] using the ReLU activation function and sigmoid mapping. The neural network automatically learns the nonlinear association between labels and type features. For example, the combination of industrial scenarios and continuous anomaly labels improves the evaluation coefficients. The model is fine-tuned and optimized using historical data, using cross-entropy as the loss function. Compared to traditional weighted summation, the model achieves significantly improved accuracy in identifying anomaly frames, while also meeting real-time inference latency requirements.

[0162] Finally, the preliminary evaluation coefficients are optimized to the final coefficients by combining scene weights and environmental adjustment factors. The scene weight table is queried based on the video frame type (for example, a weight of 1.2 for industrial production lines). The environmental factors are adjusted based on real-time network status and edge node load, and the final coefficients are calculated through weighted calculations. The final coefficients are compared with dynamic thresholds (for example, mean + 2 standard deviations) to generate the evaluation results. Scene weights enable adaptive adjustment of evaluation criteria in different scenarios, while environmental factors compensate for external influences, ensuring that the evaluation system maintains high and stable accuracy across different scenarios.

[0163] Example 2

[0164] This application also provides a real-time quality detection system based on 5G edge computing, such as Figure 2 Shown, including:

[0165] Video processing module 1, used to obtain dynamic video stream data of real-time video surveillance, and perform video frame low-latency feature extraction on the dynamic video stream data to obtain video frame feature extraction data, wherein the video frame feature extraction data includes clarity index, color distribution characteristics and brightness fluctuation characteristics;

[0166] Sequence construction module 2, used to generate a corresponding video frame quality feature matrix according to the clarity index, color distribution characteristics and brightness fluctuation characteristics, and construct a dynamic quality feature sequence based on timestamp information;

[0167] A difference analysis module 3 is configured to perform real-time difference analysis on quality feature sequences of adjacent video frames through edge nodes to obtain a quality correlation coefficient, and obtain an influence coefficient of the adjacent video frames based on the quality correlation coefficient;

[0168] A correlation graph construction module 4 is configured to construct a real-time quality correlation graph based on the influence coefficient, and obtain video frame data to be evaluated according to the real-time quality correlation graph, wherein the video frame data to be evaluated includes a frame mark sequence and video frame type information;

[0169] Evaluation and decision module 5, configured to obtain a video frame evaluation coefficient based on the frame mark sequence and the video frame type information, and determine whether the video frame evaluation coefficient is less than a preset threshold;

[0170] If it is less than or equal to, the video frame data to be evaluated is determined to be normal;

[0171] If it is greater than, the video frame data to be evaluated is determined to be abnormal.

[0172] In one embodiment, the video processing module 1 includes:

[0173] a clarity extraction unit, configured to obtain resolution, frame rate, and contrast data of consecutive frames according to a real-time video stream, and calculate a clarity index based on the resolution, frame rate, and contrast data;

[0174] A color analysis unit, configured to perform RGB and HSV color space conversion according to the color distribution characteristics of the real-time video stream, and obtain a color temperature deviation value and a saturation fluctuation range;

[0175] A feature generation unit, configured to generate a color distribution feature vector based on the color temperature deviation value and the saturation fluctuation range;

[0176] The brightness analysis unit is used to obtain an average brightness value and a maximum brightness value of a video frame according to the real-time video stream, and obtain a brightness fluctuation feature according to the average brightness value and the maximum brightness value.

[0177] Example 3

[0178] like Figure 3 As shown, embodiment 3 of the present invention provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the real-time quality detection method based on 5G edge computing provided in embodiment 1 is implemented.

[0179] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A real-time quality detection method based on 5G edge computing, characterized in that: The method comprises: Acquire dynamic video stream data of real-time video surveillance, and perform video frame low-latency feature extraction on the dynamic video stream data to obtain video frame feature extraction data, wherein the video frame feature extraction data includes a clarity index, a color distribution feature, and a brightness fluctuation feature; Generating a corresponding video frame quality feature matrix according to the clarity index, color distribution feature, and brightness fluctuation feature, and constructing a dynamic quality feature sequence based on timestamp information; In the edge node, the pre-trained difference analysis of adjacent video frames is performed to obtain the initial model parameters; Perform frame-by-frame difference calculation on the quality feature sequences of adjacent video frames to generate a feature difference matrix; Obtain the relative change rate of each feature item based on the feature difference matrix, and obtain the probability distribution of each feature item; Obtaining a characteristic anomaly entropy value according to the probability distribution; Generate a quality correlation coefficient matrix according to the initialization model parameters and characteristic anomaly entropy values; Normalize the quality correlation coefficient matrix to obtain the influence coefficients of adjacent video frames; Each video frame is represented as a node, and the influence coefficients of adjacent video frames are represented as edge weights; Construct an initial association graph based on edge weights and mark the abnormal video frame sequence position of each edge; Obtaining the associated path of each abnormal video frame according to the position of the abnormal video frame sequence; Obtain the sum of the delays for each associated path, find the path with the minimum propagation delay, and use it as the critical path; Generating a real-time quality association graph through topological structure analysis according to the critical path, and acquiring video frame data to be evaluated according to the real-time quality association graph, wherein the video frame data to be evaluated includes a frame mark sequence and video frame type information; Obtaining a video frame evaluation coefficient according to the frame mark sequence and the video frame type information, and determining whether the video frame evaluation coefficient is less than a preset threshold; If it is less than or equal to, the video frame data to be evaluated is determined to be normal; If it is greater than, the video frame data to be evaluated is determined to be abnormal.

2. The real-time quality detection method based on 5G edge computing according to claim 1 is characterized in that The step of performing low-latency feature extraction of video frames on the dynamic video stream data to obtain video frame feature extraction data includes: Obtaining resolution, frame rate, and contrast data of consecutive frames according to a real-time video stream, and calculating a clarity index based on the resolution, frame rate, and contrast data; Perform RGB and HSV color space conversion according to the color distribution characteristics of the real-time video stream, and obtain the color temperature deviation value and saturation fluctuation range; generating a color distribution feature vector based on the color temperature deviation value and the saturation fluctuation range; An average brightness value and a maximum brightness value of a video frame are obtained according to the real-time video stream, and a brightness fluctuation feature is obtained according to the average brightness value and the maximum brightness value.

3. The real-time quality detection method based on 5G edge computing according to claim 1 is characterized in that The step of generating a corresponding video frame quality feature matrix according to the clarity index, color distribution feature, and brightness fluctuation feature, and constructing a dynamic quality feature sequence based on timestamp information includes: Assign a unique timestamp to each frame of video and sort the quality feature matrix in the order of timestamps; Obtain the time interval between adjacent frames, and perform sliding window processing on the quality feature matrix based on the time interval to generate a local quality feature subsequence; Calculate the mean, variance and change rate of each local quality feature subsequence to obtain the dynamic quality feature vector; Combine the dynamic quality feature vector with the original quality feature matrix to generate a dynamic quality feature sequence; Perform discrete Fourier transform on the dynamic quality feature sequence to extract periodic quality fluctuation characteristics; The periodic quality fluctuation characteristics are integrated with the dynamic quality feature sequence to form a complete dynamic quality feature sequence.

4. The real-time quality detection method based on 5G edge computing according to claim 1 is characterized in that The step of obtaining the video frame data to be evaluated according to the real-time quality association graph includes: Obtaining video frames of a critical path according to the real-time quality association graph, and recording them as candidate frames to be evaluated; Parse the original annotation data of the candidate frame to be evaluated to obtain the frame label sequence; Acquiring a shooting scene, device type, and content features of the video frame, and generating video frame type information according to the shooting scene, device type, and content features of the video frame; Perform local sliding window processing on the dynamic quality feature sequence of the candidate frame to be evaluated to extract the time correlation features; Combine the confidence values ​​of the frame label sequence to generate a label confidence feature vector; The time correlation feature is fused with the label confidence feature vector to construct a dynamic feature correlation matrix; The video frame data to be evaluated is obtained according to the video frame type information and the dynamic feature association matrix.

5. The real-time quality detection method based on 5G edge computing according to claim 1 is characterized in that The step of obtaining a video frame evaluation coefficient according to the frame mark sequence and the video frame type information includes: Marking the frame sequence to obtain marking information, wherein the marking information includes normal marking, suspicious marking and abnormal marking; Classify and encode the video frame type information to generate a type feature vector; Construct an evaluation input matrix based on the label information and type feature vector; generating preliminary evaluation coefficients according to the evaluation input matrix; A final video frame evaluation coefficient is obtained according to the preliminary evaluation coefficient.

6. A real-time quality detection system based on 5G edge computing, characterized in that: The system comprises: A video processing module is used to obtain dynamic video stream data of real-time video surveillance and perform low-latency feature extraction of video frames on the dynamic video stream data to obtain video frame feature extraction data, wherein the video frame feature extraction data includes a clarity index, color distribution characteristics, and brightness fluctuation characteristics; A sequence construction module is used to generate a corresponding video frame quality feature matrix according to the clarity index, color distribution characteristics and brightness fluctuation characteristics, and to construct a dynamic quality feature sequence based on timestamp information; The difference analysis module is used to load pre-trained difference analysis on adjacent video frames in the edge node to obtain the initial model parameters; Perform frame-by-frame difference calculation on the quality feature sequences of adjacent video frames to generate a feature difference matrix; Obtain the relative change rate of each feature item based on the feature difference matrix, and obtain the probability distribution of each feature item; Obtaining a characteristic anomaly entropy value according to the probability distribution; Generate a quality correlation coefficient matrix according to the initialization model parameters and characteristic anomaly entropy values; Normalize the quality correlation coefficient matrix to obtain the influence coefficients of adjacent video frames; An association graph building module that represents each video frame as a node and the influence coefficients of adjacent video frames as edge weights; Construct an initial association graph based on edge weights and mark the abnormal video frame sequence position of each edge; Obtaining the associated path of each abnormal video frame according to the position of the abnormal video frame sequence; Obtain the sum of the delays for each associated path, find the path with the minimum propagation delay, and use it as the critical path; Generating a real-time quality association graph through topological structure analysis according to the critical path, and acquiring video frame data to be evaluated according to the real-time quality association graph, wherein the video frame data to be evaluated includes a frame mark sequence and video frame type information; An evaluation decision module, configured to obtain a video frame evaluation coefficient based on the frame mark sequence and the video frame type information, and determine whether the video frame evaluation coefficient is less than a preset threshold; If it is less than or equal to, the video frame data to be evaluated is determined to be normal; If it is greater than, the video frame data to be evaluated is determined to be abnormal.

7. A real-time quality detection system based on 5G edge computing according to claim 6, characterized in that: The video processing module includes: a clarity extraction unit, configured to obtain resolution, frame rate, and contrast data of consecutive frames according to a real-time video stream, and calculate a clarity index based on the resolution, frame rate, and contrast data; A color analysis unit, configured to perform RGB and HSV color space conversion according to the color distribution characteristics of the real-time video stream, and obtain a color temperature deviation value and a saturation fluctuation range; A feature generation unit, configured to generate a color distribution feature vector based on the color temperature deviation value and the saturation fluctuation range; The brightness analysis unit is used to obtain an average brightness value and a maximum brightness value of a video frame according to the real-time video stream, and obtain a brightness fluctuation feature according to the average brightness value and the maximum brightness value.

8. A computer-readable storage medium, characterized in that A program is stored thereon, and when the program is executed by the processor, the real-time quality detection method based on 5G edge computing described in any one of claims 1-5 is implemented.

Images