Image processing method and system based on artificial intelligence

Abstract

The invention provides an image processing method and system based on artificial intelligence, and relates to the field of image processing, and the method comprises the steps: determining a plurality of node clusters, image processing resource demand impact factors and enhancement impact factors according to the historical image processing data of a plurality of traffic image collection nodes; according to the image processing resource demand influence factor and the plurality of node clusters, predicting future image processing resource demands of the plurality of traffic image acquisition nodes; determining a future optimal enhancement algorithm of the plurality of traffic image acquisition nodes according to the enhancement influence factors; according to the future image processing resource requirements of the plurality of traffic image acquisition nodes and the future optimal enhancement algorithm, edge image processing nodes are allocated, and the edge image processing nodes are used for generating an enhanced traffic image according to the future optimal enhancement algorithm of the allocated traffic image acquisition nodes; and performing target detection on the enhanced traffic image by using the target detection model, so that the method has the advantage of improving the image processing efficiency and quality in an intelligent traffic scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing, in particular to an image processing method and system based on artificial intelligence. BACKGROUND

[0002] At present, intelligent transportation has become an important means to alleviate traffic pressure and improve traffic efficiency. However, the acceleration of urbanization has made traffic problems more serious, so in order to solve such problems, it is necessary to accurately and timely obtain traffic data and understand traffic information. After collecting the data of traffic information, it is particularly necessary to process the traffic image.

[0003] Image processing technology is a technology that uses computers, cameras and other digital processing technologies to operate and process images, which can conveniently extract information in images. The existing image processing method mainly includes the following steps: image enhancement, image segmentation, feature extraction, target detection and recognition, background modeling and motion target detection process. However, the existing method at least includes the following problems in the image processing process: under different conditions, the required image enhancement algorithm of the collected traffic image is different, the mismatch of the image enhancement algorithm may cause the picture distortion, the image segmentation is sensitive to interference factors, which may cause inaccurate segmentation results, the feature extraction and target recognition have high computational complexity, which is difficult to process in real time, the feedback delay, and cannot timely provide decision support for users to manage traffic.

[0004] Therefore, it is necessary to provide an image processing method and system based on artificial intelligence, which is used to improve the efficiency and quality of image processing in the intelligent transportation scene. SUMMARY

[0005] The present application provides an image processing method based on artificial intelligence, comprising: determining a plurality of node clusters of a target area according to historical image processing data of a plurality of traffic image collection nodes of the target area; determining an image processing resource demand influence factor according to the historical image processing data of the plurality of traffic image collection nodes of the target area; determining an enhancement influence factor according to the historical image processing data of the plurality of traffic image collection nodes of the target area; predicting future image processing resource demand of the plurality of traffic image collection nodes of the target area according to the image processing resource demand influence factor and the plurality of node clusters; determining future optimal enhancement algorithm of the plurality of traffic image collection nodes of the target area according to the enhancement influence factor; and allocating an edge image processing node according to the future image processing resource demand of the plurality of traffic image collection nodes of the target area and the future optimal enhancement algorithm, wherein the edge image processing node is used to perform image enhancement on a video stream according to the future optimal enhancement algorithm of the allocated traffic image collection node, generate an enhanced traffic image, and perform target detection on the enhanced traffic image using a target detection model.

[0006] Further, the historical image processing data of the traffic image collection node includes the demand of multiple image processing resources at multiple historical time points; and the multiple node clusters of the target area are determined according to the historical image processing data of the multiple traffic image collection nodes of the target area, including: for each traffic image collection node, calculating the demand fluctuation value of each image processing resource corresponding to the traffic image collection node based on the historical image processing data of the traffic image collection node; determining the key image processing resource from the multiple image processing resources according to the demand fluctuation value of each image processing resource corresponding to each traffic image collection node; calculating the demand fluctuation similarity of any two traffic image collection nodes according to the demand of the key image processing resource at multiple historical time points of any two traffic image collection nodes; and clustering the multiple traffic image collection nodes according to the demand fluctuation similarity of any two traffic image collection nodes by using an improved clustering algorithm to determine the multiple node clusters of the target area.

[0007] Further, the multiple node clusters of the target area are determined according to the demand fluctuation similarity of any two traffic image collection nodes by using an improved clustering algorithm to cluster the multiple traffic image collection nodes, including: determining the range of the number of clustering centers according to the demand fluctuation similarity of any two traffic image collection nodes; generating multiple sample clustering center schemes according to the range of the number of clustering centers; generating the clustering results corresponding to each sample clustering center scheme; generating multiple reference clustering center schemes based on the clustering results corresponding to each sample clustering center scheme by using a particle swarm algorithm; determining multiple candidate clustering centers based on the multiple reference clustering center schemes; determining multiple optimal clustering centers from the multiple candidate clustering centers based on the demand fluctuation similarity of any two candidate clustering centers; and clustering the multiple traffic image collection nodes according to the multiple optimal clustering centers and the demand fluctuation similarity of any two traffic image collection nodes to determine the multiple node clusters of the target area.

[0008] Further, the image processing resource demand influence factor is determined according to the historical image processing data of the multiple traffic image collection nodes of the target area, including: determining multiple weather influence factors and multiple road state influence factors; for each traffic image collection node, calculating the Spearman rank correlation coefficient between the weather influence factor and the image processing resource demand and the Spearman rank correlation coefficient between the road state influence factor and the image processing resource demand according to the historical image processing data of the traffic image collection node; and screening the image processing resource demand influence factor from the multiple weather influence factors and the multiple road state influence factors based on the Spearman rank correlation coefficient between the weather influence factor and the image processing resource demand and the Spearman rank correlation coefficient between the road state influence factor and the image processing resource demand corresponding to each traffic image collection node.

[0009] Further, the historical image processing data of the traffic image collection node includes optimal image enhancement algorithms of a plurality of historical time points; and the enhancement influence factor is determined according to the historical image processing data of the plurality of traffic image collection nodes of the target area, including: for each traffic image collection node, based on the historical image processing data of the traffic image collection node, calculating the Spearman rank correlation coefficient of the meteorological influence factor and the image enhancement algorithm selection and the Spearman rank correlation coefficient of the road state influence factor and the image enhancement algorithm selection; and based on the Spearman rank correlation coefficient of the meteorological influence factor and the image enhancement algorithm selection and the Spearman rank correlation coefficient of the road state influence factor and the image enhancement algorithm selection corresponding to each traffic image collection node, screening the enhancement influence factor from the plurality of meteorological influence factors and the plurality of road state influence factors.

[0010] Further, the future image processing resource demand of the plurality of traffic image collection nodes of the target area is predicted according to the image processing resource demand influence factor and the plurality of node clusters, including: for each node cluster, establishing and training a demand prediction model corresponding to the node cluster; for each traffic image collection node, acquiring first meteorological features and first road state features of the traffic image collection node according to the image processing resource demand influence factor; and for each node cluster, predicting the future image processing resource demand of the plurality of traffic image collection nodes included in the node cluster by the demand prediction model corresponding to the node cluster according to the first meteorological features, the first road state features and the current image processing resource demand of each traffic image collection node included in the node cluster.

[0011] Further, the future optimal enhancement algorithm of the plurality of traffic image collection nodes of the target area is determined, including: for each traffic image collection node, acquiring second meteorological features and second road state features of the traffic image collection node according to the image processing resource demand influence factor, and determining the future optimal enhancement algorithm of the traffic image collection node according to the second meteorological features and the second road state features of the traffic image collection node by an algorithm matching model.

[0012] Further, the edge image processing node is allocated according to the future image processing resource demand and the future optimal enhancement algorithm of the plurality of traffic image collection nodes of the target area, including: defining a Q function; sorting the plurality of traffic image collection nodes according to the future image processing resource demand of the plurality of traffic image collection nodes of the target area to generate a sorting result; and allocating the edge image processing node according to the sorting result and the future image processing resource demand and the future optimal enhancement algorithm of the plurality of traffic image collection nodes of the target area by a deep Q network according to the Q function.

[0013] Further, the Q function is related to the demand fluctuation similarity of any two traffic image acquisition nodes and the consistency of the future optimal enhancement algorithm, wherein the smaller the demand fluctuation similarity of any two traffic image acquisition nodes allocated to the same edge image processing node is, the larger the Q function value is; and / or the higher the consistency of the future optimal enhancement algorithm allocated to the same edge image processing node is, the larger the Q function value is.

[0014] The application provides an image processing system based on artificial intelligence, which is applied to the image processing method based on artificial intelligence and comprises a demand analysis module, a factor determination module, a demand prediction module, an enhancement analysis module and an edge allocation module.

[0015] Compared with the prior art, the image processing method and system based on artificial intelligence provided by the application have at least the following beneficial effects: 1. By analyzing historical data, the key factors (such as traffic flow, weather conditions, etc.) affecting image processing resource demand are determined, so that the resource demand of each node in the future can be more accurately predicted. Based on the accurate prediction result, dynamic scheduling of resources can be realized to ensure that sufficient processing resources can be provided for each node in peak periods or special situations.

[0016] 2、By selecting the optimal enhancement algorithm based on key factors such as the enhancement factor, image quality can be significantly improved, providing strong support for subsequent target detection and analysis. In different traffic scenarios and weather conditions, image features vary. By selecting the optimal enhancement algorithm, image processing needs in different scenarios can be met. High-quality images help improve the accuracy of target detection models, allowing for more accurate identification of traffic elements. Accurate target detection results can provide strong support for intelligent decision-making in traffic management, safety monitoring, and other areas, improving the overall efficiency and safety of the transportation system.

[0017] 3、Edge image processing nodes are close to data sources, allowing real-time or near-real-time processing of video streams, reducing transmission delays and processing times. By distributing processing tasks across multiple edge nodes, system reliability and fault tolerance can be improved. Even if a node fails, other nodes can continue to provide services. Edge processing reduces the amount of data that needs to be transmitted to central servers, reducing bandwidth usage and transmission costs. BRIEF DESCRIPTION OF DRAWINGS

[0018] This specification will be further illustrated in the form of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, and in these embodiments, the same numbers represent the same structures, wherein: Figure 1 is a flowchart of an artificial intelligence-based image processing method according to some embodiments of the present specification; Figure 2 is a flowchart of determining a plurality of node clusters of a target area according to some embodiments of the present specification; Figure 3 is a module diagram of an artificial intelligence-based image processing system according to some embodiments of the present specification. DETAILED DESCRIPTION

[0019] To more clearly illustrate the technical solutions of the embodiments of the present specification, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some examples or embodiments of the present specification, and for those skilled in the art, without creative labor, the present specification can also be applied to other similar scenarios according to these drawings. Unless it is obvious from the language environment or otherwise stated, the same reference numbers in the drawings represent the same structure or operation.

[0020] Figure 1 is a flowchart of an artificial intelligence-based image processing method according to some embodiments of the present specification, as Figure 1 shown, the artificial intelligence-based image processing method can include the following steps.

[0021] S101, determine a plurality of node clusters of the target region according to historical image processing data of a plurality of traffic image collection nodes of the target region.

[0022] Specifically, the historical image processing data of the traffic image collection node includes the demand for multiple image processing resources at multiple historical time points. The multiple image processing resources can at least include processor (CPU) resources, graphics processor (GPU) resources, memory resources, storage resources (such as hard disks, solid state disks, etc.), and the like. A large number of computing tasks are involved in image processing, such as image scaling, rotation, filtering, etc., which all need to be processed by CPU. GPU has a significant advantage in the field of image processing, and is used for efficiently processing complex computing tasks such as image recognition, image segmentation, feature extraction, etc. A large amount of image data and processing results need to be loaded in the image processing process, and sufficient memory resources can ensure the fast reading and writing of data and improve the processing efficiency. Storage resources are used to save historical image data, processing results and model parameters, etc., to ensure the accessibility and persistence of data.

[0023] The historical image processing data of the traffic image collection node includes the optimal image enhancement algorithm at multiple historical time points. It can be understood that under different weather conditions and road conditions, the image enhancement algorithm required for the collected traffic images is different.

[0024] For example, sunny day: the image is usually clear, but there may be overexposure or shadow areas caused by strong light. Histogram equalization can adjust the brightness distribution of the image and enhance the contrast; while adaptive histogram equalization can further enhance the local contrast while preserving the details of the image, and improve the overexposed and shadow areas.

[0025] Overcast day: the overall brightness of the image is reduced, and the contrast may be insufficient. Gamma correction can adjust the brightness of the image to make it closer to the visual habits of the human eye; Retinex algorithm can simulate the human visual system to enhance details and colors by separating the illumination and reflection components of the image.

[0026] Rainy day: there may be raindrops, rain curtains and other interference factors in the image, causing image blurring and contrast reduction. The rain removal algorithm can effectively remove raindrops and rain curtains to restore the clarity of the image; combined with histogram equalization, it can further enhance the contrast of the image.

[0027] Foggy day: the image is affected by fog, and the overall image is blurred, with significantly reduced contrast. Snowy day: the image may contain snowflakes, snow, and other interference, and the high brightness may cause overexposure. The de-fogging algorithm based on the atmospheric scattering model can restore the clarity and color of the image and improve the readability of the image.

[0028] Snowy day: The snowy day image may contain snowflakes, snow, and other interference, and the brightness is relatively high, which may cause overexposure. The snow removal algorithm based on morphological filtering can effectively remove the interference of snowflakes and snow; combined with brightness adjustment, the image brightness can be more moderate, which is convenient for subsequent analysis.

[0029] Traffic flow: When the traffic flow is large, the vehicles in the image are dense, which may increase the noise and occlusion problem. The median filter or bilateral filter can effectively remove the noise while preserving the edge information of the image; combined with contrast enhancement, the image clarity can be improved, which is convenient for identifying and analyzing vehicles.

[0030] Vehicle type: Different types of vehicles have different sizes, colors, and shapes, which have different effects on image recognition and analysis. When there are many types of vehicles, the multi-scale Retinex algorithm can enhance the details and colors of the image, and improve the recognition rate of different types of vehicles.

[0031] In some embodiments, S101 specifically comprises: For each traffic image collection node, based on the historical image processing data of the traffic image collection node, the demand fluctuation value of each image processing resource corresponding to the traffic image collection node is calculated, for example, for each image processing resource, the standard deviation of the demand of the traffic image collection node for the image processing resource at multiple historical time points is calculated as the demand fluctuation value of the traffic image collection node for the image processing resource. According to the demand fluctuation value of each image processing resource corresponding to each traffic image collection node, the key image processing resource is determined from the multiple image processing resources, for example, for each image processing resource, the mean value of the demand fluctuation value of each traffic image collection node for the image processing resource is calculated to obtain the demand fluctuation value mean value of the image processing resource, and the image processing resource whose demand fluctuation value mean value is greater than the demand fluctuation value mean value threshold is taken as the key image processing resource. According to the demand of the key image processing resource of any two traffic image collection nodes at multiple historical time points, the demand fluctuation similarity of any two traffic image collection nodes is calculated, for example, for each image processing resource, the Spearman rank correlation coefficient is calculated as the demand fluctuation similarity of the two traffic image collection nodes for the image processing resource according to the demand of the two traffic image collection nodes for the image processing resource at multiple historical time points, and then the demand fluctuation similarity of each traffic image collection node for each image processing resource is weighted and summed to obtain the demand fluctuation similarity of the two traffic image collection nodes. Through the improved clustering algorithm, the demand fluctuation similarity of any two traffic image collection nodes is calculated, and the multiple traffic image collection nodes are clustered to determine the multiple node clusters of the target region.

[0032] It can be understood that by calculating the fluctuation value (such as standard deviation) of the demand of each image processing resource corresponding to the traffic image acquisition node, the change degree of the resource demand of the node at different historical time points can be quantitatively reflected. This helps to more accurately understand the resource demand characteristics of the node and provides a basis for subsequent resource allocation and algorithm selection.

[0033] Based on the mean of the demand fluctuation value, the image processing resource with large demand fluctuation is identified as the key image processing resource. This helps to focus the limited resources and management efforts on the resources that have the greatest impact on system performance, improving resource utilization efficiency. By identifying the key image processing resource, more explicit guidance can be provided for subsequent resource allocation and algorithm selection, ensuring that core resources are fully protected.

[0034] By calculating the demand fluctuation similarity (such as using Spearman rank correlation coefficient) of any two traffic image acquisition nodes, nodes with similar resource demand fluctuation characteristics can be identified. This helps to divide nodes with similar demand into the same node cluster, facilitating subsequent resource allocation and algorithm selection. Using demand fluctuation similarity as the basis for clustering can improve the accuracy of clustering and ensure that nodes in the same node cluster have similar resource demand characteristics.

[0035] By improved clustering algorithm, according to the demand fluctuation similarity, a plurality of traffic image acquisition nodes are clustered, and a plurality of node clusters of the target area are determined. This helps to more accurately divide node clusters and improve the pertinence and effectiveness of resource allocation and algorithm selection. Through reasonable node cluster division, information support can be provided for subsequent allocation of edge image processing nodes, avoiding the allocation of multiple traffic image acquisition nodes with similar demand to the same edge image processing node, resulting in resource vacancy or resource shortage of the edge image processing node.

[0036] Figure 2 is a flowchart of determining a plurality of node clusters of a target area according to some embodiments of the present specification, as shown in Figure 2 In some embodiments, by improved clustering algorithm, according to the demand fluctuation similarity of any two traffic image acquisition nodes, a plurality of traffic image acquisition nodes are clustered, and a plurality of node clusters of the target area are determined, including: According to the demand fluctuation similarity of any two traffic image acquisition nodes, the number range of cluster centers is determined; According to the number range of cluster centers, a plurality of sample cluster center schemes are generated, wherein the sample cluster center scheme can include a plurality of traffic image acquisition nodes as cluster centers, and the number thereof is within the number range of cluster centers; generate a clustering result corresponding to each sample clustering center scheme, for example, for each clustering center number range, the traffic image collection nodes included in the clustering center number range are taken as clustering centers for clustering by a K-means clustering algorithm to generate a clustering result corresponding to the sample clustering center scheme; generate a plurality of reference clustering center schemes based on the clustering result corresponding to each sample clustering center scheme through a particle swarm algorithm; Based on the plurality of reference clustering center schemes, a plurality of candidate clustering centers are determined. Specifically, the plurality of traffic image collection nodes included in the plurality of reference clustering center schemes as clustering centers can be de-duplicated to obtain a plurality of candidate clustering centers. Based on the demand fluctuation similarity of any two candidate clustering centers, a plurality of optimal clustering centers are determined from the plurality of candidate clustering centers. According to the demand fluctuation similarity of any two traffic image collection nodes and the plurality of optimal clustering centers, the plurality of traffic image collection nodes are clustered to determine a plurality of node clusters of the target area. Specifically, the plurality of traffic image collection nodes are clustered according to the plurality of optimal clustering centers through a K-means clustering algorithm to generate a plurality of node clusters of the target area.

[0037] Specifically, a first demand fluctuation similarity sequence can be generated according to the demand fluctuation similarity of any two traffic image collection nodes, wherein each element in the first demand fluctuation similarity sequence represents the demand fluctuation similarity of the two traffic image collection nodes. The elements in the first demand fluctuation similarity sequence are sorted in ascending order to generate a second demand fluctuation similarity sequence, and the second demand fluctuation similarity sequence is subjected to Fourier transform to extract key features of the second demand fluctuation similarity sequence, such as spectral density, dominant frequency, harmonic component, and spectral bandwidth. Through the frequency domain features of the second demand fluctuation similarity sequence, the distribution and complexity of the data can be indirectly understood, thereby providing a basis for determining the clustering center number range. For example, if the spectral density is widely distributed, it indicates that the data has high complexity and requires more clustering centers; if the dominant frequency is prominent and the harmonic component is less, it indicates that the data has a simple structure and requires fewer clustering centers; the clustering center number range is determined according to the key features of the second demand fluctuation similarity sequence. For example only, the clustering center number range can be determined based on the key features of the second demand fluctuation similarity sequence through a range prediction model, wherein the range prediction model can be a convolutional neural network model.

[0038] The plurality of reference clustering center schemes can be generated based on the clustering result corresponding to each sample clustering center scheme through a particle swarm algorithm by the following process: S1011, initialize the particle swarm: each sample clustering center scheme is regarded as a particle, and each particle contains a set of clustering center nodes. A random position (i.e. a set of initial clustering center nodes) and a speed (indicating the direction and step size of the adjustment of the clustering center nodes) are assigned to each particle; S1012, calculate the fitness value of the sample clustering center scheme according to the clustering result corresponding to the sample clustering center scheme. The fitness value can be measured based on a clustering quality indicator (such as the silhouette coefficient, the Calinski-Harabasz index, etc.). The better the clustering quality, the higher the fitness value; S1013, update the speed and position of the particle according to the current position, speed and historical optimal position (individual optimal solution) of the particle, and the global optimal position (population optimal solution). The update formula usually includes an inertia weight, a cognitive coefficient and a social coefficient, which control the approximation speed of the particle to the individual optimal solution and the global optimal solution; S1014, generate a new clustering center scheme with a fitness value greater than a fitness value threshold as a candidate clustering center scheme; S1015, determine whether a predetermined number of iterations is reached, the fitness value no longer significantly improves, or the number of candidate clustering center schemes reaches a preset number. If yes, the generation of the plurality of reference clustering center schemes is completed. If no, S1013 is performed.

[0039] For each candidate clustering center, the mean of the demand fluctuation similarity of the candidate clustering center and other candidate clustering centers can be calculated. The candidate clustering centers are sorted in ascending order according to the mean of the demand fluctuation similarity of each candidate clustering center and other candidate clustering centers. The candidate clustering center with a mean of the demand fluctuation similarity less than a mean threshold is selected as an optimal clustering center, wherein the number of optimal clustering centers is within the range of the number of clustering centers.

[0040] It can be understood that the traditional clustering algorithm such as K-means clustering is easy to fall into a local optimal solution. By generating a plurality of sample clustering center schemes and clustering based on these schemes, more possibilities can be explored, thereby avoiding local optimization and improving the accuracy of the clustering result. By screening the candidate clustering centers through the particle swarm algorithm and the demand fluctuation similarity of any two candidate clustering centers, the clustering centers that can better represent the characteristics of the data can be selected, so that the clustering result is more stable and reliable.

[0041] S102, determine the image processing resource demand influence factor according to the historical image processing data of the plurality of traffic image collection nodes of the target region.

[0042] In some embodiments, S102 specifically includes: determining a plurality of weather influence factors (e.g., temperature, humidity, visibility, light intensity, etc.) and a plurality of road state influence factors (e.g., traffic volume, vehicle type proportion, etc.); For each traffic image collection node, based on the historical image processing data of the traffic image collection node, a Spearman rank correlation coefficient of the weather influence factors and the image processing resource demand and a Spearman rank correlation coefficient of the road state influence factors and the image processing resource demand are calculated; Based on the Spearman rank correlation coefficient of the weather influence factors and the image processing resource demand and the Spearman rank correlation coefficient of the road state influence factors and the image processing resource demand corresponding to each traffic image collection node, the image processing resource demand influence factors are selected from the plurality of weather influence factors and the plurality of road state influence factors.

[0043] Specifically, the weather influence factors refer to various weather conditions that affect the quality of traffic images, such as temperature, humidity, visibility, light intensity, and adverse weather conditions such as rain, snow, and fog. These factors may affect the selection of image enhancement algorithms by affecting the imaging effect of the camera.

[0044] The Spearman rank correlation coefficient is a non-parametric statistical quantity used to measure the monotonic relationship between two variables. It does not require data to follow a specific distribution, but is calculated based on the rank of the data (i.e., the position after sorting). The weather influence factors and road state influence factors with a Spearman rank correlation coefficient greater than the Spearman rank correlation coefficient threshold can be selected as the image processing resource demand influence factors.

[0045] S103, determining enhancement influence factors according to historical image processing data of a plurality of traffic image collection nodes in the target area.

[0046] In some embodiments, S103 specifically includes: For each traffic image collection node, based on the historical image processing data of the traffic image collection node, a Spearman rank correlation coefficient of the weather influence factors and the image enhancement algorithm selection and a Spearman rank correlation coefficient of the road state influence factors and the image enhancement algorithm selection are calculated; Based on the Spearman rank correlation coefficient of the weather influence factors and the image enhancement algorithm selection and the Spearman rank correlation coefficient of the road state influence factors and the image enhancement algorithm selection corresponding to each traffic image collection node, the enhancement influence factors are selected from the plurality of weather influence factors and the plurality of road state influence factors.

[0047] Specifically, the image enhancement algorithm is a variety of algorithms used to improve image quality, such as enhancement, contrast enhancement, sharpening, etc. In traffic image processing, selecting the appropriate image enhancement algorithm is crucial to improving the clarity and recognizability of the image.

[0048] The historical image processing data of each traffic image acquisition node is collected, including the record of image enhancement algorithm selection under different meteorological conditions. The meteorological influence factor and the image enhancement algorithm selection are quantitatively processed for statistical analysis. For example, the meteorological influence factor can be divided into different levels (such as high temperature, low temperature, high humidity, low humidity, etc.), and the image enhancement algorithm selection can be coded into different numerical values or categories.

[0049] The meteorological influence factor and the image enhancement algorithm selection are rank converted. That is, the values of each variable are arranged in order from small to large, and the corresponding rank is assigned. If two or more values are equal, take the average of their rank.

[0050] For each pair of meteorological influence factor and image enhancement algorithm selection data points, the difference between their ranks is calculated, and the Spearman rank correlation coefficient is calculated based on the difference between the ranks of the meteorological influence factor and the image enhancement algorithm selection corresponding to the traffic image acquisition node. The way to calculate the Spearman rank correlation coefficient of the road state influence factor and the image enhancement algorithm selection is similar to the way to calculate the Spearman rank correlation coefficient of the meteorological influence factor and the image enhancement algorithm selection, which will not be described here.

[0051] The meteorological influence factor and the road state influence factor with a Spearman rank correlation coefficient greater than the threshold value of the Spearman rank correlation coefficient can be used as an enhancement influence factor.

[0052] S104, according to the image processing resource demand influence factor and the plurality of node clusters, predicting the future image processing resource demand of the plurality of traffic image acquisition nodes in the target area.

[0053] In some embodiments, S104 specifically includes: For each node cluster, a demand prediction model corresponding to the node cluster is established and trained, wherein the demand prediction model can be a long short-term memory network model; For each traffic image acquisition node, according to the image processing resource demand influence factor, the first meteorological feature and the first road state feature of the traffic image acquisition node are obtained, wherein the first meteorological feature can include a feature value corresponding to each meteorological influence factor as an image processing resource demand influence factor, and the first road state feature can include a feature value corresponding to each road state influence factor as an image processing resource demand influence factor; For each node cluster, the future image processing resource demand of the plurality of traffic image collection nodes included in the node cluster is predicted by a demand prediction model corresponding to the node cluster according to the first meteorological feature, the first road state feature and the current image processing resource demand of each traffic image collection node included in the node cluster. Specifically, the current image processing resource demand can include the demand of each image processing resource at a plurality of time points in a current period (for example, a day, a week, etc.), and the input of the demand prediction model corresponding to the node cluster includes the first meteorological feature, the first road state feature and the current image processing resource demand of each traffic image collection node included in the node cluster and the demand fluctuation similarity of any two traffic image collection nodes included in the node cluster, and the output of the demand prediction model corresponding to the node cluster includes the future image processing resource demand of the plurality of traffic image collection nodes included in the node cluster.

[0054] It can be understood that by establishing and training a special demand prediction model for each node cluster, the demand change rule of the traffic image collection nodes in the node cluster can be more accurately captured. The long short-term memory network model is suitable for processing time series data and can effectively utilize historical information to predict future demand, thereby improving the accuracy of prediction. In the prediction process, not only the current image processing resource demand is considered, but also the meteorological feature and the road state feature and the similarity of the image processing resource demand changes of the traffic image collection nodes included in the node cluster are combined. These features as the influence factors of the image processing resource demand can more comprehensively reflect the complexity and diversity of the demand change of the traffic image collection nodes, so that the prediction result is closer to the actual situation.

[0055] S105, determining the future optimal enhancement algorithm of the plurality of traffic image collection nodes in the target area according to the enhancement influence factor.

[0056] In some embodiments, S105 specifically includes: For each traffic image collection node, the second meteorological feature and the second road state feature of the traffic image collection node are obtained according to the image processing resource demand influence factor, and the future optimal enhancement algorithm of the traffic image collection node is determined by an algorithm matching model according to the second meteorological feature and the second road state feature of the traffic image collection node, wherein the algorithm matching model can be a decision tree model, the second meteorological feature can include a feature value corresponding to each meteorological influence factor as an enhancement influence factor, and the second road state feature can include a feature value corresponding to each road state influence factor as an enhancement influence factor.

[0057] It can be understood that by considering the specific meteorological and road state characteristics of the traffic image acquisition nodes, the algorithm matching model can accurately select the most matched enhancement algorithm for the current environment. For example, in foggy or low-light environments, specific de-fogging or contrast-enhancing algorithms may be needed to improve image quality. This precise matching can avoid the use of unnecessary or inefficient enhancement algorithms, thereby optimizing the use of image processing resources.

[0058] The environment in which the traffic image acquisition nodes are located may vary due to weather, time, geographical location, and other factors. By dynamically matching the enhancement algorithm, it can ensure that the image processing system always uses the most suitable algorithm for the current environment, improving the adaptability and flexibility of the system. Selecting the most suitable enhancement algorithm for the current environment can reduce unnecessary computation and processing time, improving the efficiency of image processing. Precise matching of enhancement algorithms can more effectively improve image quality, improve image readability and accuracy, and thus provide stronger support for subsequent image analysis and processing.

[0059] S106, according to the future image processing resource demand and the future optimal enhancement algorithm of the plurality of traffic image acquisition nodes of the target area, allocate edge image processing nodes.

[0060] Among them, the edge image processing node is used to perform image enhancement on the video stream according to the future optimal enhancement algorithm of the allocated traffic image acquisition node, generate enhanced traffic images, and use a target detection model to perform target detection on the enhanced traffic images.

[0061] Specifically, the edge image processing node is a computing device deployed at the edge of the network, which is close to the data source (i.e. traffic image acquisition node) and can process video streams in real time or quasi-real time.

[0062] According to the future image processing resource demand and the future optimal enhancement algorithm of each traffic image acquisition node, appropriate edge image processing nodes are allocated to these nodes. This means that each traffic image acquisition node will have an edge image processing node to provide services for it.

[0063] After the allocation is completed, the edge image processing node will start to perform image enhancement on the video stream according to the future optimal enhancement algorithm of the allocated traffic image acquisition node. The enhanced traffic images will have higher clarity and readability, making it easier to perform target detection and analysis.

[0064] Next, the edge image processing node will use pre-trained target detection models to perform target detection on the enhanced traffic images. These models may be able to identify vehicles, pedestrians, traffic signs, and other traffic elements, and provide valuable information for subsequent traffic management, safety monitoring, and other purposes.

[0065] In some embodiments, S105 specifically includes: Define the Q function; sorting the multiple traffic image acquisition nodes according to future image processing resource requirements of the multiple traffic image acquisition nodes in the target area to generate a sorting result, for example, a traffic image acquisition node with a greater future image processing resource requirement has a higher sorting result; Through the deep Q-network, the edge image processing nodes are allocated according to the Q-function, the sorting results, the future image processing resource requirements of multiple traffic image acquisition nodes in the target area, and the future optimal enhancement algorithm.

[0066] First, multiple traffic image acquisition nodes in the target area are ranked based on their future image processing resource requirements and the similarity of demand fluctuations between any two nodes. The ranking results reflect the relative similarity of demand between nodes, providing a basis for subsequent allocation. The ranking results, the future image processing resource requirements of multiple traffic image acquisition nodes in the target area, and information about the optimal future enhancement algorithm are passed as input to the Deep Q-Network. The Deep Q-Network, through its internal neural network structure, performs complex processing and analysis on these inputs. Based on the Q function, the Deep Q-Network calculates the Q-value for each possible action (i.e., assigning an edge image processing node). The action with the highest Q-value is selected as the optimal allocation solution for the current state.

[0067] In some embodiments, the Q function is related to the demand fluctuation similarity of any two traffic image collection nodes and the consistency of the future optimal enhancement algorithm, wherein the smaller the demand fluctuation similarity of any two traffic image collection nodes assigned to the same edge image processing node, the larger the Q function value; And / or, the higher the consistency of the future optimal enhancement algorithms assigned to the same edge image processing node, the greater the Q function value.

[0068] Specifically, the Q function is defined as follows: Where s represents the current state, including information such as the future image processing resource requirements and future optimal enhancement algorithms for all traffic image acquisition nodes. a represents the action of assigning which traffic image acquisition nodes to which edge image processing nodes. In state s Next select action a Q value. N is the set of edge image processing nodes in the target area, n i ∈ N Indicates the i edge image processing nodes, n j ∈ N Indicates the ian edge image processing node. the total sum of future image processing resource demand of traffic image collection nodes assigned to the first i edge image processing node, the total sum of future image processing resource demand of traffic image collection nodes assigned to the first j edge image processing node. the total number of traffic image collection nodes assigned to the first i edge image processing node. the demand fluctuation similarity of the first i traffic image collection node and the first e traffic image collection node assigned to the first f edge image processing node, the greater the value, the higher the demand fluctuation similarity. the consistency parameter of the future optimal enhancement algorithm of the first i traffic image collection node and the first e traffic image collection node assigned to the first f edge image processing node. If the future optimal enhancement algorithm of the first e traffic image collection node and the first f traffic image collection node is consistent, then otherwise 0. α 、 β 、 γ is a weight parameter for adjusting the influence degree of different factors on the Q value, α 、 β 、 γ are all greater than 0, and α + β + γ =1.

[0069] The similarity of the nodes in resource demand is evaluated by calculating the demand fluctuation similarity of any two traffic image collection nodes and taking the average value. The smaller the demand fluctuation similarity of any two traffic image collection nodes assigned to the same edge image processing node, the greater the Q function value, indicating that these traffic image collection nodes are more suitable for being assigned to the same edge image processing node, effectively avoiding resource competition.

[0070] The similarity of the traffic image collection nodes in algorithm demand is evaluated by calculating the consistency of the future optimal enhancement algorithm assigned to the same edge image processing node. The higher the consistency, the greater the Q function value.

[0071] The resource occupancy of the edge image processing nodes is evaluated by calculating the sum of the future image processing resource requirements of the edge image processing nodes allocated to the same edge image processing node, so as to achieve load balancing of the plurality of edge image processing nodes.

[0072] It can be understood that by considering the future image processing resource requirements and requirement fluctuation similarity of the plurality of traffic image acquisition nodes, the nodes can be sorted, so that the edge image processing nodes are more reasonably allocated. This strategy can ensure efficient use of resources and avoid waste or shortage of resources. Allocating traffic image acquisition nodes with less requirement fluctuation similarity to the same edge image processing node can reduce resource competition caused by consistent requirement fluctuations. This helps to improve the stability and reliability of the entire traffic monitoring system. The higher the consistency of the future optimal enhancement algorithms allocated to the same edge image processing node, the more uniform the algorithm strategy that these nodes can use in image processing. This helps to improve the effect and consistency of image processing, so that the processed images are more in line with the needs of subsequent analysis and application. Moreover, if different image enhancement algorithms are used by traffic image acquisition nodes, algorithm switching needs to be performed when processing images of different traffic image acquisition nodes. Such switching introduces additional computational overhead and delay, affecting overall performance. Using the same image enhancement algorithm can avoid such overhead.

[0073] Through reasonable resource allocation, the delay of data transmission and processing can be reduced, and the response speed and efficiency of the system can be improved. In particular, in the traffic monitoring scene with high real-time requirements, this advantage is particularly obvious. The introduction of the Q function and the sorting result makes the resource allocation process more intelligent and automated.

[0074] Figure 3 is a schematic diagram of a module of an artificial intelligence-based image processing system according to some embodiments of the present specification, as shown in Figure 3 As shown, the artificial intelligence-based image processing system can include a requirement analysis module, a factor determination module, a requirement prediction module, an enhancement analysis module, and an edge allocation module.

[0075] The requirement analysis module is configured to determine a plurality of node clusters of a target region according to historical image processing data of a plurality of traffic image acquisition nodes of the target region. The factor determination module is configured to determine an image processing resource requirement influence factor according to the historical image processing data of the plurality of traffic image acquisition nodes of the target region. The factor determination module is further configured to determine an enhancement influence factor according to the historical image processing data of the plurality of traffic image acquisition nodes of the target region. a demand prediction module configured to predict future image processing resource demands of the plurality of traffic image collection nodes in the target area according to the image processing resource demand influence factors and the plurality of node clusters; an enhancement analysis module configured to determine future optimal enhancement algorithms of the plurality of traffic image collection nodes in the target area according to the enhancement influence factors; an edge allocation module configured to allocate edge image processing nodes according to the future image processing resource demands and the future optimal enhancement algorithms of the plurality of traffic image collection nodes in the target area, wherein the edge image processing nodes are configured to perform image enhancement on the video stream according to the future optimal enhancement algorithms of the allocated traffic image collection nodes, generate enhanced traffic images, and perform target detection on the enhanced traffic images using the target detection model.

[0076] The image processing system based on artificial intelligence can be used to perform the image processing method based on artificial intelligence, which will not be described herein.

[0077] Finally, it should be understood that the embodiments described in the specification are only used to illustrate the principles of the embodiments of the specification. Other variations can also belong to the scope of the specification. Therefore, as an example but not limitation, alternative configurations of the embodiments of the specification can be considered consistent with the teachings of the specification. Accordingly, the embodiments of the specification are not limited to the embodiments explicitly introduced and described in the specification.

Claims

1. An image processing method based on artificial intelligence, characterized in that: include: Determine multiple node clusters in the target area based on historical image processing data of multiple traffic image collection nodes in the target area; Determine the factors affecting image processing resource requirements based on historical image processing data from multiple traffic image acquisition nodes in the target area; Determine the enhancement impact factor based on historical image processing data of multiple traffic image collection nodes in the target area; Based on the image processing resource demand influencing factors and multiple node clusters, the future image processing resource demand of multiple traffic image collection nodes in the target area is predicted; Determine the optimal future enhancement algorithm for multiple traffic image acquisition nodes in the target area based on the enhancement impact factor; According to the future image processing resource requirements and future optimal enhancement algorithms of multiple traffic image acquisition nodes in the target area, edge image processing nodes are allocated, wherein the edge image processing nodes are used to perform image enhancement on the video stream according to the future optimal enhancement algorithm of the allocated traffic image acquisition nodes, generate enhanced traffic images, and use the target detection model to perform target detection on the enhanced traffic images.

2. The image processing method based on artificial intelligence according to claim 1, characterized in that: The historical image processing data of traffic image acquisition nodes includes the requirements of multiple image processing resources at multiple historical time points; Based on the historical image processing data of multiple traffic image collection nodes in the target area, multiple node clusters in the target area are determined, including: For each traffic image acquisition node, based on the historical image processing data of the traffic image acquisition node, the demand fluctuation value of each image processing resource corresponding to the traffic image acquisition node is calculated; Determine key image processing resources from a variety of image processing resources based on the demand fluctuation value of each image processing resource corresponding to each traffic image acquisition node; Calculate the demand fluctuation similarity of any two traffic image acquisition nodes based on the demand for key image processing resources at multiple historical time points; Through an improved clustering algorithm, multiple traffic image collection nodes are clustered according to the demand fluctuation similarity of any two traffic image collection nodes, and multiple node clusters in the target area are determined.

3. The image processing method based on artificial intelligence according to claim 2, characterized in that: Through the improved clustering algorithm, multiple traffic image collection nodes are clustered according to the similarity of demand fluctuations between any two traffic image collection nodes, and multiple node clusters in the target area are determined, including: According to the similarity of demand fluctuations between any two traffic image collection nodes, the range of cluster center numbers is determined; Generate multiple sample cluster center schemes according to the number range of cluster centers; Generate clustering results corresponding to each sample cluster center scheme; Through the particle swarm algorithm, multiple reference cluster center schemes are generated based on the clustering results corresponding to each sample cluster center scheme; Based on multiple reference cluster center schemes, multiple candidate cluster centers are determined; Based on the demand fluctuation similarity between any two candidate cluster centers, multiple optimal cluster centers are determined from multiple candidate cluster centers; According to multiple optimal clustering centers and the similarity of demand fluctuations between any two traffic image collection nodes, multiple traffic image collection nodes are clustered to determine multiple node clusters in the target area.

4. The image processing method based on artificial intelligence according to claim 1, characterized in that: Based on the historical image processing data of multiple traffic image acquisition nodes in the target area, factors affecting image processing resource requirements are determined, including: Determine multiple meteorological influencing factors and multiple road condition influencing factors; For each traffic image collection node, the Spearman rank correlation coefficient between the meteorological influencing factor and the image processing resource demand, as well as the Spearman rank correlation coefficient between the road state influencing factor and the image processing resource demand, are calculated based on the historical image processing data of the traffic image collection node; Based on the Spearman rank correlation coefficient between the meteorological influencing factor and the image processing resource demand corresponding to each traffic image acquisition node, as well as the Spearman rank correlation coefficient between the road state influencing factor and the image processing resource demand, the image processing resource demand influencing factor is screened from multiple meteorological influencing factors and multiple road state influencing factors.

5. The image processing method based on artificial intelligence according to claim 4, characterized in that: The historical image processing data of the traffic image collection node includes the optimal image enhancement algorithm for multiple historical time points; Based on the historical image processing data of multiple traffic image collection nodes in the target area, the enhancement influencing factors are determined, including: For each traffic image collection node, based on the historical image processing data of the traffic image collection node, the Spearman rank correlation coefficient between the meteorological influencing factor and the image enhancement algorithm selection and the Spearman rank correlation coefficient between the road state influencing factor and the image enhancement algorithm selection are calculated; Based on the Spearman rank correlation coefficient between the meteorological influencing factor corresponding to each traffic image acquisition node and the image enhancement algorithm selection, as well as the Spearman rank correlation coefficient between the road state influencing factor and the image enhancement algorithm selection, enhancement influencing factors are screened from multiple meteorological influencing factors and multiple road state influencing factors.

6. The image processing method based on artificial intelligence according to claim 4, characterized in that: Based on the image processing resource demand influencing factors and multiple node clusters, the future image processing resource demand of multiple traffic image collection nodes in the target area is predicted, including: For each node cluster, establish and train the demand prediction model corresponding to the node cluster; For each traffic image acquisition node, obtaining a first meteorological feature and a first road state feature of the traffic image acquisition node according to an image processing resource demand influencing factor; For each node cluster, the demand prediction model corresponding to the node cluster is used to predict the future image processing resource requirements of multiple traffic image acquisition nodes included in the node cluster based on the first meteorological characteristics, first road state characteristics and current image processing resource requirements of each traffic image acquisition node included in the node cluster.

7. The image processing method based on artificial intelligence according to claim 5, characterized in that: Based on the enhancement influencing factors, the optimal enhancement algorithm for multiple traffic image acquisition nodes in the target area is determined, including: For each traffic image acquisition node, the second meteorological characteristics and the second road state characteristics of the traffic image acquisition node are obtained according to the image processing resource demand influencing factors. Through the algorithm matching model, the future optimal enhancement algorithm of the traffic image acquisition node is determined according to the second meteorological characteristics and the second road state characteristics of the traffic image acquisition node.

8. The artificial intelligence-based image processing method according to any one of claims 1 to 7, characterized in that: Based on the future image processing resource requirements and future optimal enhancement algorithms of multiple traffic image acquisition nodes in the target area, edge image processing nodes are allocated, including: Define the Q function; sorting the multiple traffic image acquisition nodes according to future image processing resource requirements of the multiple traffic image acquisition nodes in the target area and generating a sorting result; Through the deep Q-network, the edge image processing nodes are allocated according to the Q-function, the sorting results, the future image processing resource requirements of multiple traffic image acquisition nodes in the target area, and the future optimal enhancement algorithm.

9. The image processing method based on artificial intelligence according to claim 8, characterized in that: The Q function is related to the demand fluctuation similarity of any two traffic image collection nodes and the consistency of the future optimal enhancement algorithm. The smaller the demand fluctuation similarity of any two traffic image collection nodes assigned to the same edge image processing node, the larger the Q function value. And / or, the higher the consistency of the future optimal enhancement algorithms assigned to the same edge image processing node, the greater the Q function value.

10. An image processing system based on artificial intelligence, characterized in that: The artificial intelligence-based image processing method according to any one of claims 1 to 9 comprises: A demand analysis module is used to determine multiple node clusters in the target area based on historical image processing data of multiple traffic image collection nodes in the target area; A factor determination module is used to determine the image processing resource demand influencing factors based on the historical image processing data of multiple traffic image acquisition nodes in the target area; The factor determination module is further used to determine the enhancement impact factor based on historical image processing data of multiple traffic image acquisition nodes in the target area; A demand prediction module is used to predict the future image processing resource requirements of multiple traffic image acquisition nodes in the target area based on the image processing resource demand influencing factors and multiple node clusters; Enhancement analysis module, used to determine the future optimal enhancement algorithm for multiple traffic image acquisition nodes in the target area based on the enhancement influencing factors; An edge allocation module is used to allocate edge image processing nodes based on the future image processing resource requirements and future optimal enhancement algorithms of multiple traffic image acquisition nodes in the target area, wherein the edge image processing nodes are used to perform image enhancement on the video stream according to the future optimal enhancement algorithm of the allocated traffic image acquisition nodes, generate enhanced traffic images, and use the target detection model to perform target detection on the enhanced traffic images.

Images