Superscreen image server performance improvement and optimization method and system based on AI

Abstract

The invention relates to the technical field of super-screen images, in particular to an AI-based super-screen image server performance improvement and optimization method and system. Firstly, feature extraction and intelligent compression are performed on a super-large image by using a CNN, and important detail information is reserved; according to the regional importance score, segmenting the image into a plurality of element image blocks and carrying out adaptive compression; generating a to-be-audited label for each element image block; the server performs joint auditing, priority ranking and multi-thread parallel loading on the element image blocks, and uses a superposition smoothing function to process spliced image data; the edge calculation module further optimizes the image quality, including color brightness equalization and super-resolution reconstruction; and finally, the bandwidth and cache management module dynamically adjusts a preloading strategy and cache content according to a user behavior analysis result. Advanced technologies in multiple fields of deep learning, image processing and user behavior analysis are combined, and efficient, high-quality and safe super-screen image service is achieved.

Description

Technical Field

[0001] The present invention relates to the field of super-screen image technology, and in particular to a method and system for improving and optimizing the performance of a super-screen image server based on AI. Background Art

[0002] With the widespread application of ultra-high-definition images in the fields of medicine, geographic information, satellite remote sensing, etc., the storage, transmission and processing of ultra-large-size images face huge challenges. Traditional image servers often encounter problems such as loading delays, insufficient bandwidth and tight computing resources when processing these ultra-large images. Especially in the ultra-screen display environment, the display effect and smoothness of the image are difficult to meet user needs, affecting the efficiency of image processing and user experience. Therefore, how to improve the processing performance of ultra-large images and solve the bottlenecks encountered in the loading and display process has become a technical problem that needs to be solved urgently.

[0003] A Chinese invention patent with publication number CN107590194A discloses a method for cutting, loading and displaying a super-large image, comprising the following steps: obtaining the maximum number of Texture objects M supported by WebGL Context; sequentially dividing the obtained super-large image into M sub-images; converting the M sub-images into M WebGLTexture objects in WebGL Context respectively; dividing the display interface of the terminal into M display areas arranged in sequence according to the maximum display resolution of the display terminal obtained; and displaying the M WebGL Texture objects in the M display areas respectively, so as to display a thumbnail or a partial magnified image corresponding to the super-large image on the display interface of the terminal. The method of the present invention can realize the rapid loading and display of super-large images, expand the maximum resolution of the displayed image in the browser WebGL programming environment, and no longer need to rely on server-side processing during image display and processing, saving network traffic and server performance, and improving the end-user experience.

[0004] However, in the field of ultra-screen image display, with the advancement of display technology and the increase in user demand, existing image servers still face the following challenges: first, the transmission efficiency of ultra-large image data is low, resulting in long loading time and poor user experience; second, image compression methods often sacrifice image quality and cannot meet the needs of high-definition display. In addition, the lack of effective data security measures makes image data vulnerable to loss or tampering during transmission. Finally, existing systems usually do not take into account the differences in user perspective and content importance, resulting in slow loading of images in key areas, affecting the overall display effect. Summary of the invention

[0005] The purpose of the present invention is to address the problems existing in the background technology and to propose an AI-based super-screen image server performance improvement and optimization method and system.

[0006] The technical solution of the present invention is a method for improving and optimizing the performance of a super-screen image server based on AI, comprising the following specific implementation steps:

[0007] S1. Use convolutional neural networks to extract features from large image data, identify key areas, use importance evaluation functions to determine the importance scores of each area, apply different compression strategies based on the importance scores of each area, then divide the compressed image into multiple meta-image blocks, and adaptively adjust the block size based on the regional importance scores, and generate audit labels for any meta-image block;

[0008] S2. Jointly audit the audited labels, calculate the priority score of each meta-image block, and dynamically sort the meta-image blocks accordingly, so that the key areas are loaded first. Combined with the user's perspective and the importance of the content, the image blocks close to the user's perspective center and rich in key details are prioritized. Then, multi-threading is used to load the image blocks in parallel, and the meta-image blocks to be displayed are seamlessly spliced.

[0009] S3, balance the color and brightness through global adjustment, eliminate the color and brightness differences in the spliced ​​image data, and improve the resolution and enhance the details of the image through the super-resolution reconstruction model to obtain an optimized spliced ​​image;

[0010] S4. By analyzing the user's historical browsing behavior and using the behavior prediction model to predict the user's access probability, the optimized spliced ​​images with high access probability are cached first, and then based on the user's current browsing location and interactive behavior, the preloaded optimized spliced ​​images are dynamically adjusted to cover the area that the user is about to view. The preload priority is calculated and sorted according to the distance of the optimized spliced ​​images and the user's operation speed, and the images that are about to be displayed are loaded first;

[0011] S5, loading the images to be displayed in sequence, and the user browses the displayed images independently;

[0012] S6. Monitor user browsing behavior and cache space in real time. If the cache space is limited, use the least recently used algorithm to dynamically update the cache, regularly clean up infrequently used cache content and update it with new content.

[0013] Preferably, the process of dividing the meta-image blocks is as follows:

[0014] S21. Use convolutional neural network to extract features from the input super large image data, identify key areas in the image through convolution layer and pooling layer, and the convolution calculation formula is:

[0015]

[0016] In the formula, y i,jRepresents the pixel value of the output image, that is, the image matrix after convolution; x i,j Represents the pixel value of the input image; w m,n represents the weight parameter in the convolution kernel, that is, the weight within the window size; M and N represent the size of the convolution kernel M×N; b represents the bias term, which is used to adjust the output after convolution;

[0017] S22. Use the importance evaluation function to determine the importance of each region in the super large image data. The importance evaluation formula is as follows:

[0018]

[0019] Where I(x,y) represents the importance score of position (x,y); A k (x, y) represents the value of the kth feature map (Activation Map) at position (x, y), that is, the activation intensity at that position; α k represents the weight of the feature map, that is, the proportion of the importance of the k-th feature in the overall image; σ represents the activation function, and the Sigmoid activation function is used in this embodiment;

[0020] S23, applying different compression strategies to compress the image according to the score of each region: for regions with higher importance, using lossless compression or low-loss compression, and for unimportant regions, using lossy compression with a high compression ratio;

[0021] S24, dividing the compressed image into a plurality of elementary image blocks {U T1 , U T2 , ..., U Ti , ..., U Tn}, the size of the block is adaptively adjusted according to the regional importance score, where U Ti represents any meta-image block;

[0022] Any image block U Ti Size:

[0023]

[0024] Where T(x,y) represents the block size of position (x,y); I(x,y) represents the importance score of position (x,y); T high and T low They represent the block sizes of high-importance and low-importance areas respectively; τ represents the importance threshold, which determines the switching point of the block size.

[0025] Preferably, the process of seamlessly splicing the elementary image blocks to be displayed is as follows:

[0026] S31, calculating the priority score of each meta-image block, and dynamically sorting the meta-image blocks accordingly, so that the key areas are loaded first, and combining the user perspective and the importance of the content, the image blocks close to the center of the user perspective and rich in key details are prioritized to ensure that the key images are loaded first;

[0027] The priority scoring formula is as follows:

[0028]

[0029] Where P i Denotes any image block U Ti Priority score; d i Represents any image block U Ti The Euclidean distance between the center store and the center of the user's current viewing angle; ε represents a small constant to prevent d i Being zero means that the denominator has no solution; w i represents the content importance weight of the meta-image block; x i and i Denotes any image block U Ti The center point coordinates; x uesr and user Indicates the coordinates of the center of the user's viewing angle;

[0030] S32, after the priority sorting is completed, the server module uses multi-threading to load the image blocks in parallel according to the multi-core computing capability of the server. Assume that the server has N use Available threads are assigned to m image blocks with the highest priority for loading. The loading time of each image block is T i , then the total parallel loading time of the server is T total for:

[0031] T total =max(T1,T2,...,T i ,...,T m );

[0032]

[0033] Where, T i represents the loading time of each meta-image block; S i represents the size of the meta-image block, i.e., the amount of data; B represents the current bandwidth rate;

[0034] S33, seamlessly splicing the meta-image blocks to be displayed, using the superposition smoothing function S(x, y) to smooth the transition area between adjacent meta-image blocks, avoiding visual discontinuity or edge blur caused by direct splicing, and completing seamless splicing of the meta-image blocks to be displayed;

[0035] The formula for the superimposed smoothing function S(x,y) is as follows:

[0036] S(x,y)=γ(x,y)·I A (x,y)+(1-γ(x,y))·I B (x,y);

[0037] Where S(x, y) represents the final rendered pixel value, that is, the smooth transition result of adjacent image blocks A and B in the splicing area; γ(x, y) represents the weight function, which determines the transition ratio of each pixel; I A (x,y) and I B (x, y) represents the pixel values ​​of adjacent image blocks A and B respectively.

[0038] Preferably, the optimization process of the optimized stitched image is as follows:

[0039] S41. Balance the color and brightness through global adjustment to eliminate the color and brightness differences in the spliced ​​image data and ensure the consistency of the image when displayed on a super-screen:

[0040]

[0041] In the formula, I adjusted (x, y) represents the image after color brightness equalization; I stitched (x, y) represents the image data after stitching; μ target , μ current Respectively represent the target brightness mean and the current brightness mean of the spliced ​​image data;

[0042] S42. The super-resolution reconstruction model SRCNN is used to improve the image resolution and enhance the details to ensure the clarity when zoomed in:

[0043] I enhanced =f SR (I adjusted (x,y));

[0044] In the formula, I enhanced represents the image after super-resolution processing; f SR () represents the super-resolution reconstruction model, which enhances the details of the image through a neural network to ensure that the details are displayed on a high-resolution screen;

[0045] S43, performing quality inspection on the optimized stitched image to obtain the optimized stitched image.

[0046] Preferably, the behavior prediction model adopts a Markov chain model structure, and its implementation method is as follows:

[0047] Assume that the user's current operation state is S tOnly depends on the previous state S t-1 , then the next operation position S can be predicted t+1 The probability is:

[0048]

[0049] In the formula, P(S t+1 |S t ) indicates that it is used to change from state S t Transfer to state S t-1 The probability of Count(S t →S t+1 ) indicates that it is used to change from state S t Transfer to state S t-1 The number of historical occurrences of j Count(S t →S j ) indicates that all slaves in state S t The total number of transfers.

[0050] Preferably, the process of generating a label to be audited for any meta-image block is as follows:

[0051] S61, randomly select the first label generation factor F g1i ∈[1, q-1], calculate the first audit label L 1i =[F g1i ]G;

[0052] Where G is a q-order base point on the elliptic curve E, and the elliptic curve E is defined in the finite field F p The elliptic curve on the curve, the point (x', y') on the curve satisfies: (y') 2 =[(x') 3 +a(x')+b]modp; p is a finite field F p The number of elements in

[0053] S62, calculate the second audit label L 2i =[H(L 1i , m i )×F g2i +F g1i ]modq;

[0054] Among them, F g2i Generate factors for the second label, satisfying F g2i ∈[1,q-1], thus defining any meta-image block U Ti Tag audit factor F ai =[F g2i ]G;m i For any image block U TiThe binary encoding form of ; H is the hash function;

[0055] S63, generating any elementary image block U Ti The audit label L i =(L 1i , L 2i ).

[0056] Preferably, the audit process of the joint audit pending audit tags is as follows:

[0057] S71. Calculate the first-level audit parameter P a1 :

[0058] S72. Calculate the secondary audit parameter P a2 :

[0059] Among them, F ai For any image block U Ti Label audit factor of m i For any image block U Ti The binary encoding form of n is the total number of received meta-image blocks;

[0060] S73, if P a1 =P a2 , then the audit of the audited labels of each meta-image block is passed, that is, each meta-image block {U T1 , U T2 , ..., U Ti , ..., U Tn} passed the integrity audit.

[0061] The technical solution of the present invention is: an AI-based super-screen image server performance improvement and optimization system, which is used in the above-mentioned AI-based super-screen image server performance improvement and optimization method, comprising:

[0062] The image optimization module is responsible for receiving the original image data and performing image compression and segmentation;

[0063] The server module is responsible for parallel processing and loading of image blocks and intelligent resource allocation;

[0064] Edge computing module, used to process part of the image data and reduce the burden on the server module;

[0065] Bandwidth and cache management module for intelligent preloading and caching of image data;

[0066] The user equipment module is used to receive and display the optimized image and support interaction with the server module.

[0067] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects:

[0068] The present invention designs a method and system for improving and optimizing the performance of a super-screen image server based on AI; first, through intelligent compression and targeted image processing, the transmission volume of super-large image data is significantly reduced, thereby improving data transmission efficiency and reducing bandwidth requirements; secondly, a deep learning model is used to extract features and enhance details of the image. This method can optimize image quality while maintaining high-resolution display, ensuring that users have a clearer and more delicate visual experience; in addition, by analyzing user behavior and predicting future browsing patterns, the system can dynamically adjust the image loading priority, achieve rapid preloading of key content, and further improve user experience; at the same time, an audit mechanism is adopted to enhance data security and ensure the integrity of the sub-image block transmission process. The present invention not only improves the performance and efficiency of the super-screen image server, but also takes into account multiple aspects of security and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a system architecture diagram of an AI-based super-screen image server performance improvement and optimization system proposed in the present invention. DETAILED DESCRIPTION

[0070] Embodiment 1, as Figure 1 As shown, the present invention proposes an AI-based super-screen image server performance improvement and optimization system, including: an image optimization module, a server module, an edge computing module, a bandwidth and cache management module, and a user equipment module.

[0071] The image optimization module is responsible for receiving the original image data and using AI technology to compress and segment the image;

[0072] The server module is responsible for parallel processing and loading of image blocks and intelligent allocation of resources;

[0073] The edge computing module is used to process part of the image data and reduce the burden on the server module;

[0074] The bandwidth and cache management module is used to intelligently preload and cache image data;

[0075] The user equipment module is used to receive and display the optimized image and supports interaction with the server module.

[0076] Embodiment 2, the present invention proposes an AI-based super-screen image server performance improvement and optimization method, which is applied to the AI-based super-screen image server performance improvement and optimization system proposed in Embodiment 1, and its specific implementation steps are as follows:

[0077] S1. The image optimization module obtains super-large image data, intelligently compresses and optimizes the super-large image data, automatically extracts important detail areas in the image and performs targeted compression, reduces the amount of data while retaining sufficient detail information. The implementation process is as follows:

[0078] S11. Use Convolutional Neural Network (CNN) to extract features from the input super large image data, and identify the key areas in the image through the Convolutional Layer and Pooling Layer. The convolution calculation formula is:

[0079]

[0080] In the formula, y i,j Represents the pixel value of the output image, that is, the image matrix after convolution; x i,j Represents the pixel value of the input image; w m,n represents the weight parameter in the convolution kernel, that is, the weight within the window size; M and N represent the size of the convolution kernel M×N; b represents the bias term, which is used to adjust the output after convolution;

[0081] S12. Use the importance evaluation function to determine the importance of each region in the super large image data. The importance evaluation formula is as follows:

[0082]

[0083] Where I(x,y) represents the importance score of position (x,y); A k (x, y) represents the value of the kth feature map (Activation Map) at position (x, y), that is, the activation intensity at that position; α k represents the weight of the feature map, that is, the proportion of the importance of the k-th feature in the overall image; σ represents the activation function, and the Sigmoid activation function is used in this embodiment;

[0084] S13. According to the score of each area, different compression strategies are applied to compress the image: for areas with higher importance, lossless compression or low-loss compression is used, while for unimportant areas, lossy compression with a high compression ratio can be used to further reduce the amount of data. The specific implementation process is as follows:

[0085] S1301, use discrete cosine transform (DCT) to perform image compression, the DCT formula is as follows:

[0086]

[0087] Where, C(u,v) represents the frequency component after DCT transformation; f(x,y) represents the pixel value of the input image at the spatial position (x,y); V represents the size of the DCT block, and 8×8 is used in this embodiment; β(u) and β(v) represent normalization coefficients, and when u=0, Otherwise, take 1;

[0088] It should be noted that the discrete cosine transform (DCT) converts the image from the spatial domain to the frequency domain, and the high-frequency part can be discarded to reduce the amount of data;

[0089] S1302, calculate the compression ratio CR, and ensure a balance between compression effect and image quality by adjusting the compression ratio of different regions. The compression ratio formula is:

[0090]

[0091] Where CR represents the compression ratio; S original Indicates the original image data size; S compressed Indicates the size of compressed image data;

[0092] S14, dividing the compressed image into a plurality of elementary image blocks {U T1 , U T2 , ..., U Ti , ..., U Tn}, the size of the block is adaptively adjusted according to the regional importance score;

[0093] Among them, U Ti represents any meta-image block;

[0094] Any image block U Ti Size:

[0095]

[0096] Where T(x,y) represents the block size of position (x,y); I(x,y) represents the importance score of position (x,y); T high and T low They represent the block sizes of high-importance and low-importance areas respectively; τ represents the importance threshold, which determines the switching point of the block size;

[0097] S15. Generate a label to be audited for each meta-image block. Ti The generation process of the audit label is as follows:

[0098] S1501, randomly select the first label generation factor F g1i ∈[1, q-1], calculate the first audit label L 1i =[F g1i ]G;

[0099] Where G is a q-order base point on the elliptic curve E, and the elliptic curve E is defined in the finite field F p The elliptic curve on the curve, the point (x', y') on the curve satisfies: (y') 2 =[(x') 3 +a(x')+b]modp; p is a finite field F p The number of elements in

[0100] S1502, calculate the second audit label L 2i =[H(L 1i , m i )×F g2i +F g1i ]mod q;

[0101] Among them, F g2i Generate factors for the second label, satisfying F g2i ∈[1,q-1], thus defining any meta-image block U Ti Tag audit factor F ai =[F g2i ]G;m i For any image block U Ti The binary encoding form of ; H is the hash function;

[0102] S1503: Generate any elementary image block U Ti The audit label L i =(L 1i , L 2i );

[0103] S16, {[(L 11 , L 21 ), U T1 ],[(L 12 , L 22 ), U T2 ],...,[(L 1i , L 2i ), U Ti ],...,[(L 1n , L 2n ), U Tn ]}Transmit to the server module.

[0104] S2, the server module receives {[(L 11 , L 21 ), U T1 ],[(L 12 , L 22 ), U T2 ],...,[(L 1i , L2i ), U Ti ],...,[(L 1n , L 2n ), U Tn ]}, and perform the following operations:

[0105] S21. Jointly audit the labels to be audited of each meta-image block. The specific implementation process is as follows:

[0106] S2101. Calculate the first-level audit parameter P a1 :

[0107] S2102. Calculate the secondary audit parameter P a2 :

[0108] Among them, F ai For any image block U Ti Label audit factor of m i For any image block U Ti The binary encoding form of n is the total number of received meta-image blocks;

[0109] S2103, if P a1 =P a2 , then the audit of the audited labels of each meta-image block is passed, that is, the server module has completely received each meta-image block {U T1 , U T2 , ..., U Ti , ..., U Tn};

[0110] S22, calculating the priority score of each meta-image block, and dynamically sorting the meta-image blocks accordingly, so that key areas are loaded first, and combining the user's perspective and the importance of the content, the image blocks close to the center of the user's perspective and rich in key details are prioritized to ensure that the key images are loaded first;

[0111] The priority scoring formula is as follows:

[0112]

[0113] Where P i Denotes any image block U Ti Priority score; d i Denotes any image block U Ti The Euclidean distance between the center store and the center of the user's current viewing angle; ε represents a small constant, which is 0.001 in this embodiment to prevent d i Being zero means that the denominator has no solution; w i represents the content importance weight of the meta-image block; x i andi Denotes any image block U Ti The center point coordinates; x uesr and user Indicates the coordinates of the center of the user's viewing angle;

[0114] S23, after the priority sorting is completed, the server module uses multi-threading to load the image blocks in parallel according to the multi-core computing capability of the server. Assume that the server has N use Available threads are assigned to m image blocks with the highest priority for loading. The loading time of each image block is T i , then the total parallel loading time of the server is T total for:

[0115] T total =max(T1,T2,...,T i ,...,T m );

[0116]

[0117] Where, T i represents the loading time of each meta-image block; S i represents the size of the meta-image block, i.e., the amount of data; B represents the current bandwidth rate;

[0118] S24, seamlessly splicing the elementary image blocks to be displayed, using the superposition smoothing function S(x,y) to smooth the transition area between adjacent elementary image blocks to avoid visual discontinuity or edge blur caused by direct splicing;

[0119] The formula for the superimposed smoothing function S(x,y) is as follows:

[0120] S(x,y)=γ(x,y)·I A (x,y)+(1-γ(x,y))·I B (x,y);

[0121] Where S(x, y) represents the final rendered pixel value, that is, the smooth transition result of adjacent image blocks A and B in the splicing area; γ(x, y) represents the weight function, which determines the transition ratio of each pixel; I A (x,y) and I B (x, y) represent the pixel values ​​of adjacent image blocks A and B respectively;

[0122] S25. Transmit the stitched image data to the edge computing module.

[0123] S3, the edge computing module further optimizes the spliced ​​image data to ensure the best visual quality and display effect. The specific implementation process is as follows:

[0124] S31. Balance the color and brightness through global adjustment to eliminate the color and brightness differences in the spliced ​​image data and ensure the consistency of the image when displayed on a super-screen:

[0125]

[0126] In the formula, I adjusted (x, y) represents the image after color brightness equalization; I stitched (x, y) represents the image data after stitching; μ target , μ current Respectively represent the target brightness mean and the current brightness mean of the spliced ​​image data;

[0127] S32. The super-resolution reconstruction model SRCNN is used to improve the image resolution and enhance the details to ensure the clarity when zoomed in:

[0128] I enhanced =f SR (I adjusted (x,y));

[0129] In the formula, I enhanced represents the image after super-resolution processing; f SR () represents the super-resolution reconstruction model, which enhances the details of the image through a neural network to ensure that the details are displayed on a high-resolution screen;

[0130] It should be noted that SRCNN (Super-Resolution Convolutional Neural Network) is a technology that uses deep learning to perform image super-resolution reconstruction. Its main goal is to convert low-resolution images into high-resolution images and restore more image details so that the image remains clear after enlargement. The SRCNN network consists of three modules: Patch extraction and representation, Non-linear mapping, and Reconstruction. These three modules correspond to three convolution operations.

[0131] S33, performing quality inspection on the optimized stitched image to ensure that there is no obvious color difference or edge discontinuity problem, and then transmitting the optimized stitched image to the bandwidth and cache management module.

[0132] S4. The bandwidth and cache management module receives the optimized stitched image and performs the following operations:

[0133] S41, obtaining the user's historical browsing behavior, and by analyzing the user's historical browsing behavior, using a behavior prediction model, by calculating the access probability of each optimized spliced ​​image, preferentially caching the optimized spliced ​​image with a higher probability, to ensure that it can be quickly loaded when the user switches;

[0134] It should be noted that the behavior prediction model adopts the Markov chain model structure, and its implementation method is as follows:

[0135] Assume that the user's current operation state is S t Only depends on the previous state S t-1 , then the next operation position S can be predicted t+1 The probability is:

[0136]

[0137] In the formula, P(S t+1 |S t ) indicates that it is used to change from state S t Transfer to state S t-1 The probability of Count(S t →S t+1 ) indicates that it is used to change from state S t Transfer to state S t-1 The number of historical occurrences of j Count(S t →S j ) indicates that all slaves in state S t The total number of transfers;

[0138] S42, based on the user's current browsing position and interactive behavior, preload the optimized stitched image that is dynamically adjusted. If the user is about to view an adjacent image area, load it in advance for seamless display. The specific implementation method is as follows:

[0139] S4201, define a preload radius R preload , which indicates the radius of the preload area outside the user's field of view:

[0140] R preload =R view +a;

[0141] In the formula, R preload Indicates the preload radius; R view Indicates the radius of the user's current field of view; a represents an expansion parameter that is dynamically adjusted according to the user's scrolling speed and direction to ensure that the preloaded area can also be covered when the user browses quickly;

[0142] S4202: For images within the preloading radius, the preloading priority is calculated according to the distance of the optimized stitched images and the user operation speed, and the priority is sorted, the images to be displayed are loaded first, and they are transmitted to the user device module in sequence;

[0143] The calculation formula for loading priority is as follows:

[0144]

[0145] Where P preload,i Indicates the loading priority of any optimized stitched image i; d i represents the distance between any optimized stitched image i and the edge of the user's visual field; v represents the user's scrolling speed; and t represents the predicted time when the user reaches the optimized stitched image i.

[0146] S5. The user device module sequentially loads the received images on the device screen and displays them to the user, who browses the displayed images.

[0147] S6. The bandwidth and cache management module monitors the user's browsing behavior in real time and monitors the cache space. If the cache space is limited, the bandwidth and cache management module regularly clears the infrequently used cache content and updates it with new content. To achieve efficient cache management, the LRU (least recently used) algorithm is used to dynamically update the cache. The implementation process is as follows:

[0148] S61, assigning a priority level p to each optimized stitched image i , and update according to the access frequency:

[0149]

[0150] In the formula, p i represents the priority of each optimized stitched image; ρ and υ represent weight coefficients, which respectively control the impact of access frequency and time decay on priority; AccessCount i represents the number of visits to any optimized spliced ​​image; t i Indicates the time difference between any optimized stitched image and the current time, that is, the difference between "last access time" and "current time";

[0151] S62: If the cache space is insufficient, select the image block with the lowest priority and clear it, to ensure that the high-priority image blocks that the user may access again can always be retained.

[0152] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A method for improving and optimizing the performance of a super-screen image server based on AI, characterized in that: The specific implementation steps include the following: S1. Use convolutional neural networks to extract features from large image data, identify key areas, use importance evaluation functions to determine the importance scores of each area, apply different compression strategies based on the importance scores of each area, then divide the compressed image into multiple meta-image blocks, and adaptively adjust the block size based on the regional importance scores, and generate audit labels for any meta-image block; S2. Jointly audit the audited labels, calculate the priority score of each meta-image block, and dynamically sort the meta-image blocks accordingly, so that the key areas are loaded first. Combined with the user's perspective and the importance of the content, the image blocks close to the user's perspective center and rich in key details are prioritized. Then, multi-threading is used to load the image blocks in parallel, and the meta-image blocks to be displayed are seamlessly spliced. S3, balance the color and brightness through global adjustment, eliminate the color and brightness differences in the spliced ​​image data, and improve the resolution and enhance the details of the image through the super-resolution reconstruction model to obtain an optimized spliced ​​image; S4. By analyzing the user's historical browsing behavior and using the behavior prediction model to predict the user's access probability, the optimized spliced ​​images with high access probability are cached first, and then based on the user's current browsing location and interactive behavior, the preloaded optimized spliced ​​images are dynamically adjusted to cover the area that the user is about to view. The preload priority is calculated and sorted according to the distance of the optimized spliced ​​images and the user's operation speed, and the images that are about to be displayed are loaded first; S5, loading the images to be displayed in sequence, and the user browses the displayed images independently; S6. Monitor user browsing behavior and cache space in real time. If the cache space is limited, use the least recently used algorithm to dynamically update the cache, regularly clean up infrequently used cache content and update it with new content.

2. According to the AI-based super-screen image server performance improvement and optimization method of claim 1, it is characterized in that: The process of dividing the meta-image block is as follows: S21. Use convolutional neural network to extract features from the input super large image data, identify key areas in the image through convolution layer and pooling layer, and the convolution calculation formula is: In the formula, y i,j Represents the pixel value of the output image, that is, the image matrix after convolution; x i,j Represents the pixel value of the input image; w m,n Represents the weight parameter in the convolution kernel, that is, the weight within the window size; M and N represent the size of the convolution kernel (M×N); b represents the bias term, which is used to adjust the output after convolution; S22. Use the importance evaluation function to determine the importance of each region in the super large image data. The importance evaluation formula is as follows: Where I(x,y) represents the importance score of position (x,y); A k (x, y) represents the value of the kth feature map (ActivationMap) at position (x, y), that is, the activation intensity at that position; α k represents the weight of the feature map, that is, the proportion of the importance of the k-th feature in the overall image; σ represents the activation function, and the Sigmoid activation function is used in this embodiment; S23, applying different compression strategies to compress the image according to the score of each region: for regions with higher importance, using lossless compression or low-loss compression, and for unimportant regions, using lossy compression with a high compression ratio; S24, dividing the compressed image into a plurality of elementary image blocks {U T1 , U T2 , ..., U Ti , ..., U Tn }, the size of the block is adaptively adjusted according to the regional importance score, where U Ti represents any meta-image block; Any image block U Ti Size: Where T(x,y) represents the block size of position (x,y); I(x,y) represents the importance score of position (x,y); T high and T low They represent the block sizes of high-importance and low-importance areas respectively; τ represents the importance threshold, which determines the switching point of the block size.

3. According to the AI-based super-screen image server performance improvement and optimization method of claim 1, it is characterized in that: The process of seamlessly stitching the meta-image blocks to be displayed is as follows: S31, calculating the priority score of each meta-image block, and dynamically sorting the meta-image blocks accordingly, so that the key areas are loaded first, and combining the user perspective and the importance of the content, the image blocks close to the center of the user perspective and rich in key details are prioritized to ensure that the key images are loaded first; The priority scoring formula is as follows: Where P i Represents any image block U Ti Priority score; d i Represents any image block U Ti The Euclidean distance between the center store and the center of the user's current viewing angle; ε represents a small constant to prevent d i Being zero results in no solution for the denominator; w i Represents the content importance weight of the meta-image block; x i and i Represents any image block U Ti The center point coordinates; x uesr and user Indicates the coordinates of the center of the user's viewing angle; S32, after the priority sorting is completed, the server module uses multi-threading to load the image blocks in parallel according to the multi-core computing capability of the server. Assume that the server has N use Available threads are assigned to m image blocks with the highest priority for loading. The loading time of each image block is T i , then the total parallel loading time of the server is T total for: T total =max(T1,T2,...,T i ,...,T m ); Where, T i represents the loading time of each meta-image block; S i represents the size of the meta-image block, i.e., the amount of data; B represents the current bandwidth rate; S33, seamlessly splicing the meta-image blocks to be displayed, using the superposition smoothing function S(x, y) to smooth the transition area between adjacent meta-image blocks, avoiding visual discontinuity or edge blur caused by direct splicing, and completing seamless splicing of the meta-image blocks to be displayed; The formula for the superimposed smoothing function S(x,y) is as follows: S(x,y)=γ(x,y)·I A (x,y)+(1-γ(x,y))·I B (x,y); Where S(x, y) represents the final rendered pixel value, that is, the smooth transition result of adjacent image blocks A and B in the splicing area; γ(x, y) represents the weight function, which determines the transition ratio of each pixel; I A (x,y) and I B (x, y) represents the pixel values ​​of adjacent image blocks A and B respectively.

4. The method for improving and optimizing the performance of a super-screen image server based on AI according to claim 1, characterized in that: The optimization process of the optimized stitched image is as follows: S41. Balance the color and brightness through global adjustment to eliminate the color and brightness differences in the spliced ​​image data and ensure the consistency of the image when displayed on a super-screen: In the formula, I adjusted (x, y) represents the image after color brightness equalization; I stitched (x, y) represents the image data after stitching; μ target , μ current Respectively represent the target brightness mean and the current brightness mean of the spliced ​​image data; S42. The super-resolution reconstruction model SRCNN is used to improve the image resolution and enhance the details to ensure the clarity when zoomed in: I enhanced =f SR (I adjusted (x,y)); In the formula, I enhanced represents the image after super-resolution processing; f SR () represents the super-resolution reconstruction model, which enhances the details of the image through a neural network to ensure that the details are displayed on a high-resolution screen; S43, performing quality inspection on the optimized stitched image to obtain the optimized stitched image.

5. According to the AI-based super-screen image server performance improvement and optimization method of claim 1, it is characterized in that: The behavior prediction model adopts the Markov chain model structure, and its implementation method is as follows: Assume that the user's current operation state is S t Only depends on the previous state S t-1 , then the next operation position S can be predicted t+1 The probability is: In the formula, P(S t+1 |S t ) indicates that it is used to change from state S t Transfer to state S t-1 The probability of Count(S t →S t+1 ) indicates that it is used to change from state S t Transfer to state S t-1 The number of historical occurrences of j Count(S t →S j ) indicates that all slaves in state S t The total number of transfers.

6. The method for improving and optimizing the performance of a super-screen image server based on AI according to claim 1, characterized in that: The generation process of the audit label for any meta-image block is as follows: S61, randomly select the first label generation factor F g1i ∈[1, q-1], calculate the first audit label L 1i =[F g1i ]G; Where G is a q-order base point on the elliptic curve E, and the elliptic curve E is defined in the finite field F p The elliptic curve on the curve, the point (x', y') on the curve satisfies: (y') 2 =[(x') 3 +a(x')+b]modp; p is a finite field F p The number of elements in S62, calculate the second audit label L 2i =[H(L 1i , m i )×F g2i +F g1i ]modq; Among them, F g2i Generate factors for the second label, satisfying F g2i ∈[1,q-1], thus defining any meta-image block U Ti Tag audit factor F ai =[F g2i ]G;m i For any image block U Ti The binary encoding form of ; H is the hash function; S63, generating any elementary image block U Ti The audit label L i =(L 1i , L 2i ).

7. The method for improving and optimizing the performance of a super-screen image server based on AI according to any one of claims 1 or 6, characterized in that: The audit process of the joint audit pending audit label is as follows: S71. Calculate the first-level audit parameter P a1 : S72. Calculate the secondary audit parameter P a2 : Among them, F ai For any image block U Ti Label audit factor of m i For any image block U Ti The binary encoding form of n is the total number of received meta-image blocks; S73, if P a1 =P a2 , then the audit of the audited labels of each meta-image block is passed, that is, each meta-image block {U T1 , U T2 , ..., U Ti , ..., U Tn } passed the integrity audit.

8. An AI-based super-screen image server performance improvement and optimization system, which is used to execute the AI-based super-screen image server performance improvement and optimization method according to any one of claims 1 to 7, characterized in that: include: The image optimization module is responsible for receiving the original image data and performing image compression and segmentation; The server module is responsible for parallel processing and loading of image blocks and intelligent resource allocation; Edge computing module, used to process part of the image data and reduce the burden on the server module; Bandwidth and cache management module for intelligent preloading and caching of image data; The user equipment module is used to receive and display the optimized image and support interaction with the server module.

Images