Image data hybrid compression and distributed storage method and device based on AI

Abstract

The invention relates to the field of artificial intelligence and image processing, and provides an AI-based image data hybrid compression and distributed storage method and device. The method comprises the following steps: acquiring to-be-processed image data, and preprocessing the to-be-processed image data; carrying out semantic recognition and region division on the preprocessed image data by adopting a lightweight Transform segmentation network to generate a plurality of semantic perception compression units; performing heterogeneous compression on the plurality of semantic perception compression units according to semantic region characteristics, and uniformly encoding the compression units into a mixed compression format to obtain compressed image data; performing fine-grained fragmentation on the compressed image data according to content importance, access frequency and a layer dependency relationship to obtain a plurality of image fragments; and dynamically storing the plurality of image segments to an edge node, a local cache or a central storage node based on a pre-trained scheduling model. In this way, on the premise that the image quality is guaranteed, efficient compression and intelligent distribution of the images can be achieved.

Description

Technical Field

[0001] The present application relates to the fields of artificial intelligence and image processing, and in particular to an AI-based hybrid compression and distributed storage method and device for image data. Background Art

[0002] With the rapid development of animation, art design, digital media, and other fields, the resolution, volume, and structural complexity of image data continue to increase. The generation and use of high-definition images, vector graphics, texture maps, and short video materials have increased significantly, driving the demand for digital image storage. In real-world work scenarios, designers and developers often need to frequently access image resources of different versions and resolutions, placing higher demands on image compression, decoding speed, and access efficiency.

[0003] However, traditional compression methods such as JPEG, JPEG2000, and HEVC mainly focus on removing pixel spatial redundancy. They lack the ability to model the semantics of image content and frequency domain features, and it is difficult to strike a balance between compression rate and image quality. In images with complex structures or high boundary clarity requirements (such as line drawings, character textures, etc.), compression is often accompanied by distortion, color shift, edge blur, and other problems.

[0004] On the other hand, the centralized storage architecture also exposes many bottlenecks in terms of concurrent access from multiple terminals, version management, and disaster recovery backup, which is not conducive to the flexible scheduling and long-term archiving of massive image materials in animation and art projects.

[0005] Therefore, there is an urgent need for a comprehensive technical solution that combines deep learning image enhancement, frequency domain analysis, adaptive compression strategy and intelligent distributed storage mechanism. Summary of the Invention

[0006] This application provides an AI-based hybrid compression and distributed storage method and device for image data, which aims to integrate frequency domain modeling, self-attention mechanism, diffusion model and lightweight transformer structure to achieve efficient image compression and intelligent distribution while ensuring image quality.

[0007] According to a first aspect of the present application, a method for hybrid compression and distributed storage of image data based on AI is provided, comprising: Acquiring image data to be processed, and preprocessing the image data to be processed; A lightweight Transformer segmentation network is used to perform semantic recognition and region segmentation on the preprocessed image data, generating multiple semantically aware compression units. Performing heterogeneous compression on the plurality of semantically-aware compression units according to semantic region characteristics, and uniformly encoding them into a hybrid compression format to obtain compressed image data; Fine-grained segmentation of the compressed image data according to content importance, access frequency, and layer dependency to obtain multiple image segments; The multiple image segments are dynamically stored in edge nodes, local caches or central storage nodes based on a pre-trained scheduling model.

[0008] As a preferred embodiment, preprocessing the image data to be processed includes: Performing frequency domain transformation on the image data to be processed to generate a low-frequency component and a high-frequency component group; Repairing the low-frequency component and high-frequency component groups through a dual-branch frequency domain processing network to generate low-frequency features and high-frequency features groups; Generate a query matrix based on the high-frequency feature group to guide low-frequency features for color correction, generate a key matrix based on the low-frequency features to guide high-frequency features for texture enhancement, generate an initial enhanced high-frequency feature group and low-frequency features, and generate initial enhanced image data through inverse discrete wavelet transform; The residuals of the initially enhanced image data and the image data to be processed are learned and optimized through a residual diffusion model, the optimized residuals are fused with the high-frequency feature group and low-frequency features of the initially enhanced data, and a high-quality output image is generated by combining an inverse discrete wavelet transform.

[0009] As a preferred embodiment, the repairing of the low-frequency component and high-frequency component groups by a dual-branch frequency domain processing network to generate low-frequency features and high-frequency features includes: The high-frequency component group is texture-restored by the wide transformer module of the dual-branch frequency domain processing network to generate a high-frequency feature group; The low-frequency components are color-restored by a space-frequency fusion module of a dual-branch frequency domain processing network to generate low-frequency features.

[0010] As a preferred embodiment, a query matrix is ​​generated based on the high-frequency feature group to guide the low-frequency features for color correction, a key matrix is ​​generated based on the low-frequency features to guide the high-frequency features for texture enhancement, and the initial enhanced high-frequency feature group and low-frequency features are generated, including: Generate a query matrix and a high-frequency generated value matrix according to the high-frequency feature group, and generate a key matrix and a low-frequency generated value matrix according to the low-frequency features; Based on the cross-frequency domain attention interaction mechanism, the edge area is focused according to the query matrix and the high-frequency generated value matrix to perform low-frequency feature color correction; constraining a high-frequency adjustment range according to the key matrix and strengthening a texture of a high-frequency feature group according to the low-frequency generated value matrix; Output the initial enhanced high-frequency feature group and low-frequency features.

[0011] As a preferred embodiment, generating a query matrix and a high-frequency generated value matrix according to the high-frequency feature group, and generating a key matrix and a low-frequency generated value matrix according to the low-frequency features, includes: The high-frequency feature group includes a horizontal high-frequency component, a vertical high-frequency component and a diagonal high-frequency component; Adding the horizontal high-frequency component, the vertical high-frequency component, and the diagonal high-frequency component element by element to generate an integrated high-frequency feature; Performing a convolution operation on the integrated high-frequency features to generate a query matrix and a high-frequency generated value matrix; A convolution operation is performed on the low-frequency features to generate a key matrix and a low-frequency generated value matrix.

[0012] As a preferred embodiment, the residuals of the initially enhanced image data and the image data to be processed are learned and optimized by a residual diffusion model, the optimized residuals are fused with the high-frequency feature group and low-frequency features of the initial enhancement, and a high-quality output image is generated by combining an inverse discrete wavelet transform, including: Inputting the initially enhanced image data and the image data to be processed into a residual diffusion model, and optimizing the residual through a reverse diffusion process; The optimized residual is combined with the initially enhanced high-frequency feature group and low-frequency feature to obtain the optimized high-frequency feature group and low-frequency feature; An inverse discrete wavelet transform is performed on the optimized high-frequency feature group and low-frequency features to generate a high-quality output image.

[0013] As a preferred embodiment, the lightweight Transformer segmentation network is used to perform semantic recognition and region division on the pre-processed image data to generate multiple semantically aware compression units, including: Construct a hybrid axial attention mechanism to decompose two-dimensional self-attention into one-dimensional attention calculations on the height axis and width axis, reducing the complexity of attention calculations; Fusion of adaptive position embedding, combining the fixed prior of absolute position encoding with the dynamic optimization capability of learnable position encoding, to enhance pixel-level positioning capability; A gating mechanism is introduced to regulate feature weights, and the contribution ratio of height axis attention to original features is dynamically adjusted through learnable weights; Based on the feature map output by the lightweight Transformer segmentation network after processing through the hybrid axial attention mechanism, adaptive position embedding and gating mechanism, semantic recognition and region segmentation are performed on the image to generate multiple semantically aware compression units; The plurality of semantically-aware compression units include a high-frequency texture area, a low-frequency texture area, and a vector image layer.

[0014] As a preferred embodiment, heterogeneous compression is performed on the plurality of semantically-aware compression units according to semantic region characteristics, including: Adaptive residual coding based on a deep convolutional network is used for the high-frequency texture area to compress and optimize the local residual features, and a perceptual loss function is introduced to maintain subjective visual consistency; The low-frequency background area is compressed using JPEG2000 or HEVC standards to achieve a better trade-off between compression rate and decoding complexity; A structure-preserving compression strategy is adopted for the vector layer, including path simplification, path reconstruction under topological constraints and point set regularization, so as to reduce the encoding cost while ensuring geometric accuracy.

[0015] As a preferred embodiment, dynamically storing the multiple image segments to an edge node, a local cache, or a central storage node based on a pre-trained scheduling model includes: Extracting a multidimensional feature vector for each image segment; Input the multidimensional feature vector into a pre-trained scheduling model and output a storage decision probability distribution; Based on the relationship between the decision probability distribution and the preset probability threshold, the corresponding image segments are dynamically stored in the edge node, the local cache or the central storage node.

[0016] According to a second aspect of the present application, there is provided an AI-based hybrid compression and distributed storage device for image data, comprising: An image acquisition and preprocessing module is used to acquire image data to be processed and preprocess the image data to be processed; The semantic recognition and segmentation module uses a lightweight Transformer segmentation network to perform semantic recognition and region segmentation on pre-processed image data, generating multiple semantically aware compression units; An image compression module, configured to perform heterogeneous compression on the plurality of semantically-aware compression units according to semantic region characteristics, and uniformly encode them into a hybrid compression format to obtain compressed image data; A compressed image segmentation module is used to segment the compressed image data into fine-grained segments according to content importance, access frequency and layer dependency to obtain multiple image segments; An image segment scheduling module is used to dynamically store the multiple image segments to an edge node, a local cache or a central storage node based on a pre-trained scheduling model.

[0017] According to a third aspect of the present application, an electronic device, at least one processor, and a memory communicatively connected to the at least one processor are provided; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the method described in any one of the above items.

[0018] According to a fourth aspect of the present application, a computer-readable storage medium is provided, wherein computer instructions are used to enable the computer to execute the method according to any one of the above items.

[0019] Compared with the prior art, this application achieves the following beneficial effects: (1) This application proposes a dual-branch frequency domain processing structure and designs a dual-branch network consisting of a wide transformer block (WTB) and a space-frequency fusion block (SFFB). The WTB focuses on high-frequency texture modeling, and the SFFB integrates spatial convolution and frequency amplitude enhancement to achieve the coordinated restoration of color and structural features, significantly improving the image detail reconstruction capability. (2) This application constructs an attention mechanism of "high-frequency guidance + low-frequency constraint" to enable the edge texture area to actively guide the color correction area, and the low-frequency structure limits the texture enhancement range, thereby achieving dynamic coordination between color consistency and texture sharpness, and overcoming the artifacts and color difference problems caused by high- and low-frequency separation.

[0020] (3) This application proposes a residual diffusion model (FRDAM) based on U-Net. It uses the diffusion-anti-diffusion process to simulate the degradation and restoration path for the noise and blurred details remaining after the initial enhancement of the image. The residual details are estimated and reconstructed through the U-Net network, and finally fused back to the original image frequency band. Compared with traditional denoising algorithms, it effectively improves the texture fidelity and denoising ability of the image.

[0021] (4) This application designs a lightweight Transformer network architecture that supports dynamic width adjustment. Combining the convolutional blocks in the encoder with the hybrid axial attention (HAA) mechanism, the decoder adopts skip connections and interpolation restoration strategies, making the network well adaptable to images of different resolutions and semantic complexities, significantly reducing the number of parameters and computational costs. (5) The lightweight Transformer network architecture adopted in this application integrates absolute position encoding (APE) and learnable position encoding (LPE), introduces spatial pixel-level inductive bias, and dynamically optimizes frequency expression based on data learning, effectively solving the limitations of relative position encoding (RPE) in fine positioning and improving the accuracy of boundary area segmentation and spatial consistency.

[0022] It should be understood that the contents described in the Summary of the Invention are not intended to limit the key or important features of the embodiments of the present application, nor are they intended to limit the scope of the present application. Other features of the present application will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The above and other features, advantages and aspects of the embodiments of the present application will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are provided for a better understanding of the present invention and do not constitute a limitation of the present application. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which: Figure 1 A flowchart of the AI-based hybrid compression and distributed storage method for image data according to an embodiment of the present application is shown; Figure 2 A flowchart of preprocessing image data to be processed according to an embodiment of the present application is shown; Figure 3 A block diagram of an AI-based hybrid compression and distributed storage device for image data according to an embodiment of the present application is shown; Figure 4 A schematic diagram of an exemplary electronic device according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0024] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0025] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0026] like Figure 1 FIG. 1 is a flow chart of an AI-based hybrid compression and distributed storage method for image data of the present application. The method 100 includes: S110: Acquire image data to be processed, and pre-process the image data to be processed.

[0027] In some embodiments, in order to achieve high-fidelity compression and intelligent enhancement of image data, the input image data to be processed needs to be preprocessed, including: S111: performing frequency domain transformation on the image data to be processed to generate a low-frequency component and a high-frequency component group; S112: Repairing the low-frequency component and high-frequency component groups through a dual-branch frequency domain processing network to generate low-frequency features and high-frequency features groups; S113: Generate a query matrix based on the high-frequency feature group to guide low-frequency features for color correction, generate a key matrix based on the low-frequency features to guide high-frequency features for texture enhancement, generate an initial enhanced high-frequency feature group and low-frequency features, and generate initial enhanced image data through inverse discrete wavelet transform; S114: performing learning optimization on the residuals of the initial enhanced image data and the image data to be processed through a residual diffusion model, fusing the optimized residuals with the high-frequency feature group and low-frequency features of the initial enhancement, and combining with inverse discrete wavelet transform to generate a high-quality output image.

[0028] In the above step S111, the image data to be processed is subjected to frequency domain transformation to generate a low-frequency component and a high-frequency component group, which specifically includes the following steps: First, the image is mapped from pixel space to frequency domain, and frequency band separation is used to reduce redundant information and provide structural support for subsequent frequency band-specific optimization.

[0029] Use Haar wavelet to image data Perform discrete wavelet transform (DWT) to decompose it into low-frequency components and high-frequency components: ; in, :For images The result of a two-dimensional discrete wavelet transform; Represents the low-frequency component of the image, mainly containing color information; It is a high-frequency component group, including texture and edge details.

[0030] The image is then converted from the spatial domain to the frequency domain through Fourier transform (FFT) to obtain amplitude and phase information, so that operations such as image enhancement and compression can be more accurately optimized based on frequency domain characteristics (such as energy distribution and phase constraints): ; ; in, : The frequency domain position of the image after Fourier transform The complex coefficients at : The amplitude of the frequency component indicates the intensity of the frequency in the image. : The phase of the frequency component represents the spatial positioning characteristics of the frequency in the image. : The real part of : The imaginary part of .

[0031] In the above step S112, in order to achieve the coordinated restoration of color and structural features and significantly improve the image detail reconstruction capability, this embodiment designs a dual-branch frequency domain processing network consisting of a wide transformer block (WTB) and a space-frequency fusion block (SFFB), wherein the WTB focuses on high-frequency texture modeling, and the SFFB integrates spatial convolution and frequency amplitude enhancement. Specifically, the wide transformer module of the dual-branch frequency domain processing network performs texture restoration on the high-frequency component group to generate a high-frequency feature group, including the following steps: Step 1121: Build the Wide Transformer Block (WTB) For high frequency component group The WTB module uses its self-attention (SA) to model the global dependency between textures and channel attention (CA) to strengthen the response of key details. The specific process is as follows: Input processing: ; in, : Output feature map of the previous WTB layer; : Normalization operation, used to stabilize the training process; : Linear projection weight matrix, used to map input features to a uniform dimensional space; : Dimensional reorganization weights, used to further transform the shape or channel distribution of feature representation; : Split operation, which divides the transformed features into Q (query vector), K (key vector), V (value vector), and L (input for channel attention); Q: Query, the query vector in the attention mechanism, used to ask "questions", K: Key, the key vector in the attention mechanism, used to match the query, V: Value, the value vector in the attention mechanism, used to generate the final attention weighted result, L: The input feature vector used by the channel attention module (CA), which retains all channel information for calculating weights.

[0032] Step 1122: The self-attention mechanism (SA) is used to capture global texture, and the channel attention module (CA) filters effective details to update the features: ; in, : Self-attention mechanism (SA), which uses Q, K, and V to calculate the dependencies between different positions in the feature map, emphasizing the connection between long-distance textures in the image; : Channel attention module, used to extract more significant texture / detail information in the channel dimension; : The intermediate updated feature map of the current layer WTB, which integrates the self-attention results, channel attention responses and residual information.

[0033] Step 1123: Optimize the output via a feed-forward neural network (FFN): ; in, :For the intermediate feature map Perform normalization processing; : Feedforward neural network (FFN), a two-layer fully connected network for further feature transformation and nonlinear modeling; : The final output feature map of the current layer serves as the input of the next layer, fusing the global and local frequency domain features and completing feature reconstruction.

[0034] Step 1123: Construct a space-frequency fusion block (SFFB). Use the space-frequency fusion module (SFFB) to perform color restoration on the low-frequency component and generate low-frequency features. Specifically: The SFFB module fuses the spatial domain convolution features with the frequency domain amplitude enhancement, and the final feature map after the fusion of spatial and frequency domains , which contains both spatial structure information and global color information in the frequency domain: ; in, : The spatial feature map extracted by multi-scale convolution is mainly used to restore the structure and color information of the image, especially for the low-frequency components ; : Indicates pixel-by-pixel fusion operation, keeping the channels and sizes consistent and enhancing feature expression capabilities; : The feature map reconstructed from the frequency domain reflects the optimized frequency domain amplitude and phase information, and the expression is: ; in, : Inverse Fast Fourier Transform, converting the frequency domain representation back to the spatial domain; : The optimized amplitude information is mainly used to repair the brightness and overall color of the image, which comes from Frequency domain representation of ; : The optimized phase information retains the edge and structural details of the image, also by Extracted from and : The amplitude and phase are adjusted and enhanced through a dedicated module for high-quality reconstruction.

[0035] In step S113, this embodiment designs an attention mechanism with high-frequency guidance and low-frequency constraint to achieve dynamic coordination between color consistency and texture sharpness. This mechanism overcomes the problems of traditional methods that cause disconnection between color and texture information (e.g., blurred edges after color correction) and artifacts and chromatic aberration caused by the separation of high and low frequencies due to independent processing of high and low frequencies. This mechanism achieves coordinated enhancement of color and texture. Specifically, the following steps are included: Step S1131: High frequency guides low frequency: Generate a query matrix using the high-frequency feature group generated in step S112 and high-frequency generated value matrix Based on the cross-frequency domain attention interaction mechanism, we focus on the edge areas that need color correction and optimize the low-frequency component color: ; ; in, : High-frequency feature groups, corresponding to horizontal high-frequency, vertical high-frequency and diagonal high-frequency components, mainly contain texture and edge information; + (addition operation): element-by-element summation of the three high-frequency feature groups to integrate full-frequency edge / texture information; Convolution operation, used for feature compression, channel transformation and generating attention input; : The query matrix is ​​generated by the high-frequency feature group, representing the "focus points" at the edge texture of the image, guiding the color adjustment to focus on these areas; : A matrix of high-frequency generated values ​​used to preserve texture information in the attention output.

[0036] Step S1132: Low frequency constrains high frequency: According to the key matrix generated in step S112 and the low-frequency generated value matrix , according to the bond matrix Constrain the high frequency adjustment range and generate values ​​based on the low frequency Improved texture output to ensure that texture enhancement does not destroy color consistency.

[0037] ; ; in, : Low-frequency features, mainly including the color and structure information of the image; : Key matrix, used to constrain the weight calculation of the attention mechanism to ensure that high-frequency adjustments do not destroy the overall color consistency; : Low-frequency feature generation value matrix, used to add low-frequency color constraint information to the final fusion output.

[0038] Step S1133: The attention weighted output is: ; ; in, : query matrix With the low frequency generated value matrix The inner product operation measures the matching degree between high-frequency “focus points” and low-frequency “guiding content”; : The dimension of the Key vector, used to scale the inner product value to prevent gradient instability caused by excessive values; : Convert matching weights into probability distribution for weighted summation; : The final texture output retains high-frequency information and is guided and regulated by low-frequency information to avoid over-sharpening; : The final color output is guided by low frequency to complete color constraints while responding to high frequency requirements; : Indicates linear transformation operations (such as channel mapping), which are used to further adjust the output shape.

[0039] The above attention mechanism realizes the collaborative optimization of "color-texture", corrects the color and sharpens the edges simultaneously in the edge area of ​​the image, realizes the dynamic coordination of color consistency and texture sharpness, and overcomes the artifacts and chromatic aberration problems caused by high- and low-frequency separation.

[0040] In step S114, in order to effectively improve the texture fidelity and denoising capability of the image, image reconstruction is performed based on residual diffusion. Specifically, for the texture and color image data optimized in step S113, the diffusion-anti-diffusion process is used to simulate the degradation and restoration path, and the residual details are estimated and reconstructed through the U-Net network, and finally fused back to the original image band. Specifically, the following steps are performed: S1141: Forward diffusion of the residual error of "true value-initial enhancement" Add noise gradually to simulate the degradation process: , ; ; ; in, : The residual between the initial enhanced image and the true image (the original image data to be processed), which contains noise, blur, missing details and other information to be optimized; : ground truth image; : Initial enhanced image; : No. The diffusion state of the step represents the degraded version of the current residual after adding noise; : No. The diffusion state of the step; : The retention factor in the diffusion process, which controls the proportion of residual information of the previous step retained in the current state; 1- : Attenuation factor, which controls the intensity of the added noise; : No. Gaussian noise added in the first step; : The mean is 0 and the covariance is Gaussian distribution is used to simulate random noise in the degradation process of natural images; : Given the initial residual The conditional probability distribution of The noise image distribution of the step; :forward Cumulative retention factor of the step; : The identity matrix, indicating that the covariance is a diagonal matrix (no correlation).

[0041] S1142: Backward diffusion estimates noise through U-Net and recovers residual and fused with the initial enhancement result.

[0042] ; in, : The mean of the back diffusion prediction, indicating the The step residual plot predicts the clear residual of the previous step; : Noise adjustment coefficient, which controls the intensity of noise removal; :Use the noise estimation network output learned by U-Net to try to estimate The noise part in : Multi-scale contextual features to assist estimation (e.g., low-frequency or local information).

[0043] S1143: Frequency components after residual fusion optimization: ; ; in, : High-frequency diffusion reconstruction module, which performs residual enhancement on high-frequency components and enhances fine structures (such as image texture); : Low-frequency diffusion reconstruction module, used to optimize color consistency and avoid oversaturation or color cast; : High-frequency residual estimation result output by U-Net; : Low-frequency residual estimation result output by U-Net; :The first enhanced High-frequency component images; : Initially enhanced low-frequency component image; : The optimized high-frequency residual information is used to enhance texture and edges; : The optimized low-frequency residual information is used to correct low-frequency attributes such as color and brightness.

[0044] S1144: Reconstruct the image through inverse transform. The IDWT operation reassembles the frequency domain features into a spatial domain image through inverse wavelet transform, restoring details and color consistency: ; in, : Inverse discrete wavelet transform, which restores the frequency domain features back to the spatial image; : The final high-quality image output integrates the details and color information after preliminary enhancement and diffusion optimization; : The high-frequency part after fusion (edge, texture); : The low-frequency part after fusion (structure, color).

[0045] S120: A lightweight Transformer segmentation network is used to perform semantic recognition and region segmentation on the preprocessed image data to generate multiple semantically aware compression units.

[0046] In some embodiments, in order to achieve efficient image compression and subsequent semantic structure analysis, the embodiments of the present application introduce a lightweight and adjustable transformer architecture - the slimmed-down transformer, to adapt to the requirements of different resolutions and different semantic complexities in image segmentation tasks. The network is based on the encoder-decoder structure, combined with the hybrid axial attention mechanism (HAA), adaptive position embedding (ADPE) and gating mechanism, to achieve the number of parameters is only 1 / 10 of the traditional U-Net, which significantly reduces the computational burden while also improving the expression ability of boundaries and texture structures.

[0047] A lightweight Transformer segmentation network is used to perform semantic recognition and region segmentation on the preprocessed image data to generate multiple semantically aware compression units. The specific steps include: Step 121: Build Hybrid Axis Attention Mechanism (HAA) HAA decomposes the traditional two-dimensional self-attention into two one-dimensional channels: height-axis attention and width-axis attention, to reduce computational complexity and improve the ability to model long-distance dependencies.

[0048] Among them, the height axis attention establishes the relationship between pixels along the vertical direction of the image, and its attention calculation formula is as follows: ; in, :Location The attention-weighted output features in the height axis direction, : The height of the image, that is, the number of rows of the feature map, :Location The query vector at , :Location The key vector (key) at , traversing along the height axis, :Location The value vector at ; (value) : Attention weight, measuring position and relevance; : Batch normalization operation to improve training stability; : Perform an average operation along the channel dimension to compress feature redundancy and improve generalization ability.

[0049] After modeling the height axis, the width-axis attention propagates feature information horizontally to achieve horizontal context aggregation and further enhance global perception capabilities.

[0050] The HAA mechanism reduces the attention computation complexity from the traditional Reduce to , effectively supports real-time segmentation of large-size images, where is the height of the image, that is, the number of columns of the feature map.

[0051] Step 121: Building Adaptive Positional Embedding (ADPE) To compensate for the inductive bias problem of Transformer in pixel-level positioning, ADPE combines fixed absolute position encoding (APE) and learnable position encoding (LPE) to enhance position perception. The fusion formula is as follows: ; in, : Feature representation after fusion position encoding; : Original feature map (features after attention mechanism or convolution processing); : A learnable position encoding matrix automatically optimizes spatial position perception capabilities through training; : Fixed sinusoidal position coding matrix, introducing spatial prior information; : Element-by-element addition operation, fusing features and position bias information.

[0052] Compared with using only APE or relative position encoding (RPE), the solution provided in this embodiment has higher positioning accuracy in tasks such as image boundaries, and significantly enhances segmentation accuracy and spatial consistency.

[0053] Step 123: Build the Gating Mechanism To further improve the robustness and expression efficiency of the model in small sample scenarios, this embodiment introduces a gating mechanism to perform weighted control on the axial attention features.

[0054] Specifically, learnable gating weights are introduced Dynamically adjust the contribution of height axis and width axis attention. The formula is: ; in, : The fused output features contain the weighted information of axial attention and original features; : A learnable gating factor automatically adjusts the importance of attention features based on training data; : Features processed by the hybrid axial attention module, representing the contextual information after HAA modeling; : The feature map after fusion position embedding is used as the input of HAA; : Original features The channel average of is used to preserve the overall information and enhance stability.

[0055] The above gating factors It dynamically adjusts with the data scale, effectively suppressing background noise and redundant patterns. The Dice score can be improved by 4.2%, amplifying the feature expression ability on large data sets, and converging to a lower value under small sample conditions to prevent overfitting, balancing accuracy and efficiency.

[0056] Step S124: The image is semantically recognized and region-divided using the feature map output after processing by the hybrid axial attention mechanism, adaptive position embedding, and gating mechanism to generate multiple semantically aware compression units.

[0057] S130: Performing heterogeneous compression on the plurality of semantically-aware compression units according to semantic region characteristics, and uniformly encoding them into a hybrid compression format to obtain compressed image data.

[0058] In some embodiments, heterogeneous compression strategies are adopted to achieve an optimal balance between high compression rate and visual fidelity according to the different characteristics of image content.

[0059] First, based on the image segmentation results of step S120, the system performs semantic recognition and regional division on the image content. By identifying important content areas (such as people, textures, and edge structures) and relatively less important areas (such as background and flat areas) in the image, the entire image is divided into multiple semantically aware compression units, where the multiple semantically aware compression units include high-frequency texture areas, low-frequency texture areas, and vector layers.

[0060] Among them, high-frequency texture areas (such as complex patterns, human faces, etc.) are usually more sensitive to visual quality. Therefore, adaptive residual coding or perceptual compression technology can be used. This method introduces a deep convolutional network to perform compression optimization on local residual features, while maintaining subjective visual consistency through the perceptual loss function.

[0061] Since the low-frequency texture area has simple texture and more redundant information, standard compression algorithms (such as JPEG2000 and HEVC) can be used to obtain good compression effects, achieving a better trade-off between compression rate and decoding complexity.

[0062] Vector layers (such as SVG graphics and illustration outlines) use a structure-preserving compression strategy, including path simplification, path reconstruction under topological constraints, and point set regularization, thereby reducing encoding costs while ensuring geometric accuracy.

[0063] Ultimately, all image areas are marked as AI compression blocks or traditional compression blocks based on content characteristics and uniformly encoded into a hybrid compression format. This format is not only compatible with existing image processing tools, but also facilitates differentiated decoding and loading in subsequent distributed storage architectures, which can improve the overall performance of the system.

[0064] S140: Fine-grained segmentation of the compressed image data according to content importance, access frequency, and layer dependency to obtain multiple image segments.

[0065] In some embodiments, to meet the long-term storage, high-speed access, and high-availability requirements of large-scale, high-definition image materials in industries such as fine arts, animation, and gaming, this embodiment builds an intelligent distributed storage system based on artificial intelligence scheduling. Specifically, it includes the following two aspects: (a) At the infrastructure level, the system is built on a distributed file system. Optional technologies include HDFS (Hadoop-Distributed-File-System), Ceph, or optimized systems based on object storage protocols to support horizontal expansion, metadata management, and multi-replica disaster recovery capabilities.

[0066] (b) Intelligent sharding mechanism: The compressed image data is sharded in a fine-grained manner based on "content importance", "access frequency", and "layer dependencies". For example, core layers and high-access frequency areas are split separately to facilitate fast retrieval and local loading.

[0067] S150: Dynamically storing the multiple image segments to an edge node, a local cache, or a central storage node based on a pre-trained scheduling model.

[0068] In some embodiments, in order to achieve dynamic storage of multiple image fragments, a multidimensional feature vector can be extracted for each image fragment; the multidimensional feature vector is then input into a pre-trained scheduling model, and a storage decision probability distribution is output; further, based on the relationship between the above decision probability distribution and a preset probability threshold, the corresponding image fragment is dynamically stored in an edge node, a local cache or a central storage node.

[0069] The above-mentioned dynamic storage scheduling strategy for image segments also includes the following: Historical access logs and upstream and downstream task dependencies can be used to train scheduling models (such as schedulers based on reinforcement learning or attention mechanisms), thereby dynamically deciding whether each image fragment should be stored in an edge node (to improve response speed), in the user workstation local cache (to improve interaction efficiency), or in a central node (to ensure long-term storage and consistency).

[0070] The system automatically adjusts the data heat level to ensure that frequently accessed data resides on fast nodes first, and low-frequency data is transferred to low-cost cold storage.

[0071] Each distributed node has a built-in intelligent cache module that automatically records and updates the "most recently used image blocks" and periodically synchronizes data with remote copies through an asynchronous mechanism, enhancing the system's data recovery capabilities in the event of node failure or network interruption.

[0072] The system also supports layer version control and rollback. Each layer acts as an independent data unit, allowing for quick rollback to historical versions, satisfying the trial-and-error and traceability requirements of creative design and version evolution. It also supports partial image loading, meaning users only need to load the image area or layer they want to edit, eliminating the need to load the entire image, improving editing response speed.

[0073] According to the above embodiments of the present application, the following technical effects are achieved: (1) This application proposes a dual-branch frequency domain processing structure and designs a dual-branch network consisting of a wide transformer block (WTB) and a space-frequency fusion block (SFFB). The WTB focuses on high-frequency texture modeling, and the SFFB integrates spatial convolution and frequency amplitude enhancement to achieve the coordinated restoration of color and structural features, significantly improving the image detail reconstruction capability. (2) This application constructs an attention mechanism of "high-frequency guidance + low-frequency constraint" to enable the edge texture area to actively guide the color correction area, and the low-frequency structure limits the texture enhancement range, thereby achieving dynamic coordination between color consistency and texture sharpness, and overcoming the artifacts and color difference problems caused by high- and low-frequency separation.

[0074] (3) This application proposes a residual diffusion model (FRDAM) based on U-Net. It uses the diffusion-anti-diffusion process to simulate the degradation and restoration path for the noise and blurred details remaining after the initial enhancement of the image. The residual details are estimated and reconstructed through the U-Net network, and finally fused back to the original image frequency band. Compared with traditional denoising algorithms, it effectively improves the texture fidelity and denoising ability of the image.

[0075] (4) This application designs a lightweight Transformer network architecture that supports dynamic width adjustment. Combining the convolutional blocks in the encoder with the hybrid axial attention (HAA) mechanism, the decoder adopts skip connections and interpolation restoration strategies, making the network well adaptable to images of different resolutions and semantic complexities, significantly reducing the number of parameters and computational costs. (5) The lightweight Transformer network architecture adopted in this application integrates absolute position encoding (APE) and learnable position encoding (LPE), introduces spatial pixel-level inductive bias, and dynamically optimizes frequency expression based on data learning, effectively solving the limitations of relative position encoding (RPE) in fine positioning and improving the accuracy of boundary area segmentation and spatial consistency.

[0076] It should be noted that, for the aforementioned method embodiments, for simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all optional embodiments, and the actions and modules involved are not necessarily required by this application.

[0077] The above is an introduction to the method embodiment. The following is a device embodiment to further illustrate the solution described in this application.

[0078] Figure 2 FIG1 shows a block diagram of an AI-based image data hybrid compression and distributed storage device according to an embodiment of the present application. Figure 2 As shown, the apparatus 200 includes: The image acquisition and preprocessing module 210 is used to acquire image data to be processed and preprocess the image data to be processed; Semantic recognition and segmentation module 220, for performing semantic recognition and region segmentation on the pre-processed image data using a lightweight Transformer segmentation network to generate multiple semantically aware compression units; An image compression module 230 is configured to perform heterogeneous compression on the plurality of semantically-aware compression units according to semantic region characteristics, and uniformly encode them into a hybrid compression format to obtain compressed image data; The compressed image segmentation module 240 is configured to segment the compressed image data into fine-grained segments according to content importance, access frequency, and layer dependency to obtain multiple image segments; The image segment scheduling module 250 is configured to dynamically store the plurality of image segments to an edge node, a local cache, or a central storage node based on a pre-trained scheduling model.

[0079] In the technical solution of this application, the acquisition, storage and application of user personal information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0080] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0081] In the technical solution of this application, the acquisition, storage and application of user personal information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0082] According to an embodiment of the present application, the present application also provides an electronic device, a readable storage medium and a computer program product.

[0083] Figure 4 A schematic block diagram of an electronic device 400 that can be used to implement an embodiment of the present application is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present application described and / or required herein.

[0084] The electronic device 400 includes a computing unit 401, which can perform various appropriate actions and processes according to a computer program stored in a ROM 402 or a computer program loaded from a storage unit 408 into a RAM 403. The RAM 403 may also store various programs and data required for the operation of the electronic device 400. The computing unit 401, the ROM 402, and the RAM 403 are connected to each other via a bus 404. An I / O interface 405 is also connected to the bus 404.

[0085] Multiple components in the electronic device 400 are connected to the I / O interface 405, including an input unit 406, such as a keyboard, a mouse, etc.; an output unit 407, such as various types of displays, speakers, etc.; a storage unit 408, such as a magnetic disk, an optical disk, etc.; and a communication unit 409, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 409 allows the electronic device 400 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0086] Computing unit 401 may be any general-purpose and / or specialized processing component with processing and computing capabilities. Some examples of computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Computing unit 401 performs the various methods and processes described above, such as method 100. For example, in some embodiments, method 400 may be implemented as a computer software program tangibly embodied in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program may be loaded and / or installed onto electronic device 400 via ROM 402 and / or communication unit 409. When the computer program is loaded into RAM 403 and executed by computing unit 401, one or more steps of method 100 described above may be performed. Alternatively, in other embodiments, computing unit 401 may be configured to perform method 100 in any other suitable manner (e.g., via firmware).

[0087] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system-on-chip systems (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system comprising at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0088] The program code for implementing the methods of the present application can be written in any combination of one or more programming languages. Such program code can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the program code is executed by the processor or controller, the functions / operations specified in the flow charts and / or block diagrams are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0089] In the context of this application, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or apparatus. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of machine-readable storage media may include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), optical fibers, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0090] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0091] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0092] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact through a communication network. The client-server relationship arises through computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, a server in a distributed system, or a server integrated with a blockchain.

[0093] In the technical solution of this application, the acquisition, storage and application of user personal information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0094] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this application can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solution of this application can be achieved. This is not limited herein.

[0095] The above specific embodiments do not constitute a limitation on the scope of protection of this application. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application shall be included within the scope of protection of this application.

Claims

1. An AI-based hybrid compression and distributed storage method for image data, characterized in that: include: Acquiring image data to be processed, and preprocessing the image data to be processed; A lightweight Transformer segmentation network is used to perform semantic recognition and region segmentation on the preprocessed image data, generating multiple semantically aware compression units. Performing heterogeneous compression on the plurality of semantically-aware compression units according to semantic region characteristics, and uniformly encoding them into a hybrid compression format to obtain compressed image data; Fine-grained segmentation of the compressed image data according to content importance, access frequency, and layer dependency to obtain multiple image segments; The multiple image segments are dynamically stored in edge nodes, local caches or central storage nodes based on a pre-trained scheduling model.

2. The AI-based hybrid compression and distributed storage method for image data according to claim 1, characterized in that: Preprocessing the image data to be processed includes: Performing frequency domain transformation on the image data to be processed to generate a low-frequency component and a high-frequency component group; Repairing the low-frequency component and high-frequency component groups through a dual-branch frequency domain processing network to generate low-frequency features and high-frequency features groups; Generate a query matrix based on the high-frequency feature group to guide low-frequency features for color correction, generate a key matrix based on the low-frequency features to guide high-frequency features for texture enhancement, generate an initial enhanced high-frequency feature group and low-frequency features, and generate initial enhanced image data through inverse discrete wavelet transform; The residuals of the initially enhanced image data and the image data to be processed are learned and optimized through a residual diffusion model, the optimized residuals are fused with the high-frequency feature group and low-frequency features of the initially enhanced data, and a high-quality output image is generated by combining an inverse discrete wavelet transform.

3. The AI-based hybrid compression and distributed storage method for image data according to claim 2, characterized in that: The repairing of the low-frequency component and the high-frequency component group by a dual-branch frequency domain processing network to generate a low-frequency feature and a high-frequency feature group includes: The high-frequency component group is texture-restored by the wide transformer module of the dual-branch frequency domain processing network to generate a high-frequency feature group; The low-frequency components are color-restored by a space-frequency fusion module of a dual-branch frequency domain processing network to generate low-frequency features.

4. The AI-based hybrid compression and distributed storage method for image data according to claim 2, characterized in that: Generating a query matrix based on the high-frequency feature group to guide low-frequency features for color correction, generating a key matrix based on the low-frequency features to guide high-frequency features for texture enhancement, and generating initially enhanced high-frequency feature groups and low-frequency features, including: Generate a query matrix and a high-frequency generated value matrix according to the high-frequency feature group, and generate a key matrix and a low-frequency generated value matrix according to the low-frequency features; Based on the cross-frequency domain attention interaction mechanism, the edge area is focused according to the query matrix and the high-frequency generated value matrix to perform low-frequency feature color correction; constraining a high-frequency adjustment range according to the key matrix and strengthening a texture of a high-frequency feature group according to the low-frequency generated value matrix; Output the initial enhanced high-frequency feature group and low-frequency features.

5. The AI-based hybrid compression and distributed storage method for image data according to claim 3, characterized in that: Generating a query matrix and a high-frequency generated value matrix according to the high-frequency feature group, and generating a key matrix and a low-frequency generated value matrix according to the low-frequency features, including: The high-frequency feature group includes a horizontal high-frequency component, a vertical high-frequency component and a diagonal high-frequency component; Adding the horizontal high-frequency component, the vertical high-frequency component, and the diagonal high-frequency component element by element to generate an integrated high-frequency feature; Performing a convolution operation on the integrated high-frequency features to generate a query matrix and a high-frequency generated value matrix; A convolution operation is performed on the low-frequency features to generate a key matrix and a low-frequency generated value matrix.

6. The AI-based hybrid compression and distributed storage method for image data according to claim 2, characterized in that: The residuals of the initially enhanced image data and the image data to be processed are optimized through a residual diffusion model, the optimized residuals are fused with the high-frequency feature group and low-frequency feature of the initially enhanced image data, and an inverse discrete wavelet transform is combined to generate a high-quality output image, including: Inputting the initially enhanced image data and the image data to be processed into a residual diffusion model, and optimizing the residual through a reverse diffusion process; The optimized residual is combined with the initially enhanced high-frequency feature group and low-frequency feature to obtain the optimized high-frequency feature group and low-frequency feature; An inverse discrete wavelet transform is performed on the optimized high-frequency feature group and low-frequency features to generate a high-quality output image.

7. The AI-based hybrid compression and distributed storage method for image data according to claim 1, characterized in that: The lightweight Transformer segmentation network is used to perform semantic recognition and region division on the pre-processed image data to generate multiple semantically aware compression units, including: Construct a hybrid axial attention mechanism to decompose two-dimensional self-attention into one-dimensional attention calculations on the height axis and width axis, reducing the complexity of attention calculations; Fusion of adaptive position embedding, combining the fixed prior of absolute position encoding with the dynamic optimization capability of learnable position encoding, to enhance pixel-level positioning capability; A gating mechanism is introduced to regulate feature weights, and the contribution ratio of height axis attention to original features is dynamically adjusted through learnable weights; Based on the feature map output by the lightweight Transformer segmentation network after processing through the hybrid axial attention mechanism, adaptive position embedding and gating mechanism, semantic recognition and region segmentation are performed on the image to generate multiple semantically aware compression units; The plurality of semantically-aware compression units include a high-frequency texture area, a low-frequency texture area, and a vector image layer.

8. The AI-based hybrid compression and distributed storage method for image data according to claim 7, characterized in that: Performing heterogeneous compression on the plurality of semantically aware compression units according to semantic region characteristics, including: Adaptive residual coding based on a deep convolutional network is used for the high-frequency texture area to compress and optimize the local residual features, and a perceptual loss function is introduced to maintain subjective visual consistency; The low-frequency background area is compressed using JPEG2000 or HEVC standards to achieve a better trade-off between compression rate and decoding complexity; A structure-preserving compression strategy is adopted for the vector layer, including path simplification, path reconstruction under topological constraints and point set regularization, so as to reduce the encoding cost while ensuring geometric accuracy.

9. The AI-based hybrid compression and distributed storage method for image data according to claim 1, characterized in that: Dynamically storing the plurality of image segments to an edge node, a local cache, or a central storage node based on a pre-trained scheduling model includes: Extracting a multi-dimensional feature vector for each image segment; Input the multidimensional feature vector into a pre-trained scheduling model and output a storage decision probability distribution; Based on the relationship between the decision probability distribution and the preset probability threshold, the corresponding image segments are dynamically stored in the edge node, the local cache or the central storage node.

10. An AI-based hybrid compression and distributed storage device for image data, characterized in that: include: An image acquisition and preprocessing module is used to acquire image data to be processed and preprocess the image data to be processed; The semantic recognition and segmentation module uses a lightweight Transformer segmentation network to perform semantic recognition and region segmentation on pre-processed image data, generating multiple semantically aware compression units; An image compression module, configured to perform heterogeneous compression on the plurality of semantically-aware compression units according to semantic region characteristics, and uniformly encode them into a hybrid compression format to obtain compressed image data; A compressed image segmentation module is used to segment the compressed image data into fine-grained segments according to content importance, access frequency and layer dependency to obtain multiple image segments; An image segment scheduling module is used to dynamically store the multiple image segments to an edge node, a local cache or a central storage node based on a pre-trained scheduling model.

Images