Flexible AMOLED image compression driving method based on convolutional neural network

Abstract

The invention discloses a flexible AMOLED (Active Matrix Organic Light Emitting Diode) image compression driving method based on a convolutional neural network. The method comprises the following steps: S1, inputting original image data into a multi-level feature extraction convolutional neural network module; s2, performing feature extraction on the original image data to obtain a feature map set containing global features and local detail features; s3, constructing a color mapping conversion convolutional neural network module, inputting the feature map set into the color mapping conversion convolutional neural network module, and obtaining feature data adaptive to display of the flexible AMOLED screen; s4, constructing a compression decision convolutional neural network module, and performing adaptive compression on the feature data; and S5, inputting the compressed feature data into a flexible AMOLED driving module, and realizing real-time display of an image on a flexible AMOLED screen by loading image data of different areas in a time-sharing manner. Accurate color space conversion and pixel rearrangement are achieved, screen display is perfectly adapted, the phenomena of color distortion and pixel arrangement disorder are effectively avoided, and the image display quality is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing and display, in particular to a flexible AMOLED image compression driving method based on a convolutional neural network. BACKGROUND

[0002] With the rapid development of display technology, flexible AMOLED screens have been widely used in smart phones, wearable devices, foldable screen electronic devices and other fields due to their advantages of thinness, bendability, high contrast ratio and fast response speed. However, in practical applications, flexible AMOLED screens face challenges in data transmission and processing. Due to their high resolution and rich color display requirements, image data is huge, and the storage and transmission bandwidth requirements are extremely high. If these data cannot be effectively processed, not only will the hardware cost increase, but also display delays, lag and other problems will occur, which will seriously affect the user experience. At the same time, in order to achieve long battery life and efficient operation of the device, it is necessary to reduce the energy consumption of data processing and transmission while ensuring image display quality. Therefore, it is urgent to develop an efficient flexible AMOLED image compression driving method, which can solve the bottleneck problem of data transmission and processing, improve the overall performance of the device and meet the growing mobile display requirements.

[0003] In terms of image compression, traditional flexible AMOLED image driving methods often use general compression algorithms such as JPEG. These algorithms achieve compression by quantizing, transforming and encoding the pixel data of the image. The advantage is that the algorithm is mature and widely used, and to some extent it can reduce the amount of image data and reduce storage and transmission costs. For example, in some simple display scenarios where image quality requirements are not high, JPEG compression can quickly process images, and the file size after compression is significantly reduced.

[0004] However, the traditional method has many shortcomings. On the one hand, general compression algorithms do not fully consider the special pixel arrangement and color response characteristics of flexible AMOLED screens. The pixel arrangement of flexible AMOLED screens is different from that of traditional liquid crystal screens, such as the common diamond arrangement, which makes it difficult for general algorithms to accurately operate on pixels when processing images of such screens, and is prone to color distortion and detail loss. On the other hand, traditional algorithms have difficulty in balancing compression rate and image quality during compression. When pursuing high compression rate, image quality will decrease significantly, with problems such as blocking effect and blurring; while to ensure high image quality, the compression rate cannot meet the actual demand, resulting in heavy data transmission and storage burden. In addition, the traditional fixed driving strategy cannot be dynamically adjusted according to image content and screen state, and cannot fully utilize the advantages of flexible AMOLED screens, affecting the display effect and device performance.

[0005] The prior art has begun to attempt to combine machine learning techniques in the field of flexible AMOLED image compression driving. For example, some deep learning-based image compression methods use neural network models to compress and decompress images. Compared with traditional algorithms, these methods can learn the complex features of images and improve the compression effect and image quality to some extent. They can automatically extract key information from images, reduce unnecessary data storage and transmission, and perform better in handling complex images.

[0006] However, the deep learning-based compression methods of the prior art often have high computational complexity and require powerful computing hardware support, which increases the cost and energy consumption of devices. For resource-constrained mobile devices, excessive computing requirements result in severe device heating and shortened battery life. Secondly, while general image features are considered, the unique display characteristics of flexible AMOLED screens, such as pixel arrangement and color response optimization, are not sufficiently deep. In key areas such as color gamut mapping and pixel rearrangement, existing technical solutions cannot fully accurately adapt to flexible AMOLED screens, resulting in color deviation and edge blur when displaying some specific colors and detailed images. In addition, the existing technology does not do a good job in the coordination of image compression and driving, and the compressed images have compatibility problems during driving and display, affecting the real-time display effect and stability of the images. SUMMARY

[0007] Based on the above technical problems, the present application discloses a flexible AMOLED image compression driving method based on a convolutional neural network, comprising:

[0008] S1, obtaining original image data to be displayed, and inputting the original image data into a pre-trained multi-level feature extraction convolutional neural network module;

[0009] S2, performing feature extraction on the original image data through the multi-level feature extraction convolutional neural network module to obtain a feature map set containing global features and local detail features;

[0010] S3, constructing a color mapping conversion convolutional neural network module according to the pixel arrangement mode and color response characteristics of the flexible AMOLED screen, inputting the feature map set into the color mapping conversion convolutional neural network module, and performing color space conversion and pixel rearrangement processing on the feature map set through the color mapping conversion convolutional neural network module to obtain feature data suitable for display on the flexible AMOLED screen;

[0011] S4, a compressed decision convolutional neural network module is constructed, which is used for adaptively compressing feature data according to the feature importance distribution of the feature data of the adaptive flexible AMOLED screen, and in the compression process, a dynamic threshold control strategy is adopted to dynamically adjust the compression threshold according to the frequency distribution of the feature data, so as to balance the compression rate and the image quality.

[0012] S5, the compressed feature data is input into a flexible AMOLED driving module, and a time division multiplexing driving strategy is adopted to realize real-time display of images on the flexible AMOLED screen by loading image data of different regions in time division.

[0013] Preferably, the multi-level feature extraction convolutional neural network module in S1 comprises at least three convolutional layers with different kernel sizes, wherein the first convolutional layer adopts a large size kernel for global feature extraction, and the subsequent convolutional layers adopt small size kernels for local detailed feature extraction.

[0014] Preferably, the multi-level feature extraction convolutional neural network module further comprises a feature enhancement unit, which is composed of a spatial attention submodule and a channel attention submodule in cascade, wherein the spatial attention submodule first performs global average pooling and global maximum pooling operations on the feature map output by the convolutional layer, concatenates the two pooling results in the channel dimension, and inputs them into a multi-layer perceptron containing one hidden layer, to generate a spatial attention weight matrix through a Sigmoid activation function, and then performs weight adjustment on the spatial dimension of the feature map through element multiplication; the channel attention submodule performs global average pooling and global maximum pooling on the feature map in the width and height dimensions, respectively, inputs the pooling results into a shared multi-layer perceptron network, processes them through a ReLU activation function and a full connection layer, and then generates a channel attention weight vector through a Sigmoid activation function, to perform weight scaling on the channel dimension of the feature map in a broadcast form; the weighted feature maps output by the two submodules are added element by element, and the enhanced feature map is input into the next convolutional layer.

[0015] Preferably, the color mapping conversion convolutional neural network module in S3 is constructed as follows:

[0016] The color response matrix is constructed by collecting the RGB sub-pixel brightness response data of the flexible AMOLED screen under different gray scales where m is the number of gray scales, and n is the number of measurements; the color conversion function is constructed according to the color response matrix, and the formula is: c (x) = W x + b, where x is the pixel vector of the input feature map, is the conversion weight matrix obtained by least squares optimization according to the color response matrix R, is a bias vector; the color conversion function is embedded into a fully connected layer after a convolution layer, and a color space conversion is performed on a feature map set, wherein a pixel rearrangement process is performed by constructing an index matrix Pixels are re-indexed according to an arrangement mode of the flexible AMOLED screen to realize pixel rearrangement, and feature data suitable for display of the flexible AMOLED screen is obtained, wherein h and w are respectively a height and a width of the feature map.

[0017] Preferably, the color mapping conversion convolutional neural network module realizes efficient color space conversion and accurate pixel rearrangement of the feature map through a double-branch structure, a first branch adopts a U-Net structure with a hollow convolution to perform multi-scale feature fusion on the feature map, and local and global color information of the image is captured; a second branch is a pixel arrangement prediction branch, and the offset Δp of the pixel arrangement of the flexible AMOLED screen is predicted through a structure composed of a convolution layer and an up-sampling layer. ij , wherein i and j are pixel coordinates; the feature map output by the first branch is fused with the offset predicted by the second branch, and the formula y ij = Pixel rearrangement is realized, wherein x r is an input feature map, y is output feature data, and g is a bilinear interpolation function.

[0018] Preferably, the feature data suitable for display of the flexible AMOLED screen obtained in S3 is input into an adaptive color gamut mapping layer after color space conversion and pixel rearrangement processing, a color distribution histogram H(c) of the feature map is calculated, and is compared with a pre-stored color gamut boundary B(c) of the flexible AMOLED screen to generate a color gamut mapping function , wherein c represents a color channel; for a color component exceeding the color gamut of the screen, a gradient-preserving nonlinear compression algorithm C(c) = c·exp(-α·(c-B(c)) 2 ) is used for processing, wherein α is a compression coefficient dynamically adjusted according to a screen color gamut coverage degree; the feature map after pixel rearrangement is geometrically corrected through a pixel arrangement compensation module to obtain a corrected feature map; feature fusion and dimension reduction are performed on the corrected feature map through a depth separable convolution layer to output adaptive feature data.

[0019] Preferably, the corrected feature map is input into a spatial attention-guided depth convolution layer through the depth separable convolution layer for feature fusion and dimension reduction, spatial feature extraction is performed on each channel, a channel attention weight vector is generated through global average pooling, the feature map output by the depth convolution is multiplied with the channel attention weight vector channel by channel to obtain a spatial enhancement feature map F S, the channel number of the spatial enhanced feature map is reduced to the target channel number through a point convolution layer, the point convolution layer adopts a grouping convolution strategy, the input channels are divided into target groups, each group independently performs convolution to generate an intermediate feature map F r , through a dynamic routing mechanism, the weight matrix W g (γ) of the grouping convolution is adaptively adjusted according to the color gamut characteristic parameter γ of the flexible AMOLED screen, and the output adaptive feature data is formula: Optimal color representation is achieved on the RGB three channels, wherein represents the intermediate feature map of the lth group, and G is the target group number.

[0020] Preferably, the S4 compression decision convolutional neural network module comprises a double-path structure, the main path is composed of cascaded residual convolution blocks and gated recurrent units, the residual convolution blocks extract local and global context information of the feature data through a dilated convolution, and the gated recurrent units perform time series modeling on the feature sequence to capture long-distance dependency; the auxiliary path adopts a self-attention mechanism module, calculates an attention weight matrix of the feature data, and realizes multi-scale feature interaction through a multi-head attention mechanism; the output features of the two paths are weighted and summed through a feature fusion layer, the weights are dynamically generated by display parameters of the flexible AMOLED screen, and an importance score vector of the output feature data is adaptively compressed.

[0021] Preferably, the S4 dynamic threshold control strategy is specifically: the importance score vector of the feature data output by the compression decision convolutional neural network module is estimated for probability density, the display resolution R and the color depth D of the flexible AMOLED screen are obtained, and the dynamic compression threshold T is calculated, formula: wherein β is a threshold adjustment factor, formula: P(x) is a frequency distribution value, x is a score vector value, μ is a score vector mean, and σ is a standard deviation, the feature data with an importance score lower than the threshold T is quantitatively compressed, the data higher than the threshold is losslessly compressed, and the threshold adjustment factor β is iteratively updated to realize dynamic balance between compression rate and image quality.

[0022] Preferably, the time division multiplexing driving strategy of the flexible AMOLED driving module in S5 is specifically: the compressed feature data is divided into a plurality of non-overlapping region blocks O according to spatial positions, in each display period, the region block data is loaded in sequence according to a priority sequence, full resolution is loaded for high-priority regions, and down-sampling is loaded for low-priority regions, time domain multiplexing is performed on different region block data of adjacent display periods through time interleaving technology, formula: wherein D(t) is the driving data at time t, w k is a weight coefficient, and p k(t)For the kth loading area of the tth period, K is the number of areas loaded per period, and the real-time display of the image on the flexible AMOLED screen is realized.

[0023] Compared with the prior art, the technical scheme of the application has the following technical effects:

[0024] The application constructs a color mapping conversion convolutional neural network module, performs color space conversion and pixel rearrangement processing according to the pixel arrangement mode and color response characteristics of the flexible AMOLED screen, in the color space conversion aspect, a color response matrix is constructed by using the collected RGB sub-pixel brightness response data of the screen under different gray scales, and then a color conversion function is obtained, accurate color conversion is realized, and it is ensured that the image presents accurate and bright colors on the flexible AMOLED screen; in the pixel rearrangement process, the pixels are re-indexed according to the unique arrangement mode of the screen by constructing an index matrix, and the offset predicted by the pixel arrangement prediction branch in the double-branch structure is fused and rearranged, so that the image details are clearer and the edges are sharper, the color components exceeding the color gamut of the screen are processed by using a gradient-preserving nonlinear compression algorithm through the adaptive color gamut mapping layer, the color performance is further optimized, the color distortion is reduced, the feature data is output after feature fusion and dimension reduction through the depth separable convolutional layer, and optimal color representation is realized on the three channels of RGB, and the display quality of the image on the flexible AMOLED screen is comprehensively improved, so that the user has a more realistic and delicate visual experience.

[0025] The compression decision convolutional neural network module adopts a double-path structure and a dynamic threshold control strategy, and adaptively compresses the feature data suitable for the display of the flexible AMOLED screen, in the double-path structure, the residual convolutional block of the main path cooperates with the gated recurrent unit to extract local and global context information and perform time series modeling, and the self-attention mechanism module of the auxiliary path realizes multi-scale feature interaction, and the combination of the two can accurately judge the importance of the feature data, and the dynamic threshold control strategy calculates a threshold adjustment factor according to the display resolution and color depth of the flexible AMOLED screen, estimates the probability density of the importance score vector of the feature data to dynamically adjust the compression threshold. The data with an importance score lower than the threshold is quantized and compressed, and the data with an importance score higher than the threshold is losslessly compressed, and the threshold adjustment factor is iteratively updated to realize the dynamic balance between the compression rate and the image quality. This intelligent compression method greatly reduces the data amount, reduces the data transmission bandwidth demand and storage cost, improves the data processing efficiency, and makes the device more efficient and economical in data transmission and storage.

[0026] The flexible AMOLED driving module adopts a time division multiplexing driving strategy, effectively realizes real-time display of an image on the flexible AMOLED screen, divides compressed feature data into a plurality of non-overlapping area blocks according to spatial positions, determines a priority sequence according to importance and motion of area content, in each display period, loads area block data in priority order, adopts full resolution loading for high priority areas to ensure clarity of key content, and adopts down-sampling loading for low priority areas to reduce data volume on the premise of not affecting overall visual effect, through time interleaving technology, different area block data of adjacent display periods are time domain multiplexed, and driving data at each moment is accurately calculated by using a formula, the driving strategy fully considers the display characteristics of the flexible AMOLED screen, reduces the conflict and delay of data loading, improves the refresh rate of the screen, so that the image can be displayed smoothly and in real time, effectively optimizes the display performance of the device, avoids image lag, ghosting and other problems, improves the interactive experience of the user when operating the device, and meets the needs of high dynamic image display scenes such as video playing and game playing.

[0027] The above description is only a summary of the technical scheme of the present application, in order to more clearly understand the technical means of the present application, so as to implement according to the content of the description, and in order to make the above and other purposes, characteristics and advantages of the present application more obvious and easy to understand, the following will be described in detail with the preferred embodiment of the present application and the accompanying drawings.

[0028] According to the detailed description of the specific embodiments of the present application in the following combined with the drawings, those skilled in the art will be more clear about the above and other purposes, advantages and characteristics of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical scheme in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description, obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creating creative labor. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, each element or part is not necessarily drawn according to the actual proportion.

[0030] Figure 1 A flow chart of a flexible AMOLED image compression driving method based on convolutional neural network according to the present application;

[0031] Figure 2 A double branch structure diagram of the color mapping conversion convolutional neural network module according to the present application;

[0032] Figure 3A dual-path structure diagram of the compression decision convolutional neural network module of the present application;

[0033] Figure 4 A compression ratio comparison diagram of the flexible AMOLED image compression driving method of the present application and the prior art;

[0034] Figure 5 A PSNR comparison diagram of the flexible AMOLED image compression driving method of the present application and the prior art;

[0035] Figure 6 A SSIM comparison diagram of the flexible AMOLED image compression driving method of the present application and the prior art. DETAILED DESCRIPTION

[0036] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments of the present application. In the following description, specific details such as specific configurations and components are provided only to help a comprehensive understanding of the embodiments of the present application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. In addition, in order to be clear and concise, the description of known functions and structures is omitted in the embodiments.

[0037] It should be understood that the "one embodiment" or "the embodiment" mentioned throughout the specification means that the specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present application. Therefore, "one embodiment" or "the embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner.

[0038] In addition, reference numerals and / or letters can be repeated in different examples in the present application. Such repetition is for the purpose of simplification and clarity, and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0039] The term "and / or" herein is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that there are three cases of A alone, B alone and A and B together. The term "and" herein is a description of another association relationship of the associated objects, which means that there can be two relationships, for example, A and B can mean that there are two cases of A alone and A and B together. In addition, the character " / " herein generally represents that the associated objects before and after are in an "or" relationship.

[0040] The term "at least one" is merely used to describe a corresponding relationship of associated objects, and indicates that three relationships can exist, for example, at least one of A and B can indicate that A exists alone, A and B exist together, and B exists alone.

[0041] It should also be noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion.

[0042] Embodiment 1

[0043] This embodiment mainly describes a flexible AMOLED image compression driving method based on convolutional neural network, as shown in Figure 1 Specifically, it includes:

[0044] S1, obtaining original image data to be displayed, and inputting the original image data into a pre-trained multi-level feature extraction convolutional neural network module;

[0045] S2, performing feature extraction on the original image data by the multi-level feature extraction convolutional neural network module to obtain a feature map set containing global features and local detail features;

[0046] S3, constructing a color mapping conversion convolutional neural network module according to the pixel arrangement mode and color response characteristics of the flexible AMOLED screen, inputting the feature map set into the color mapping conversion convolutional neural network module, and performing color space conversion and pixel rearrangement processing on the feature map set by the color mapping conversion convolutional neural network module to obtain feature data suitable for display on the flexible AMOLED screen;

[0047] S4, constructing a compression decision convolutional neural network module, which performs adaptive compression on the feature data according to the feature importance distribution of the feature data suitable for display on the flexible AMOLED screen. In the compression process, a dynamic threshold control strategy is adopted to dynamically adjust the compression threshold according to the frequency distribution of the feature data, so as to balance the compression rate and image quality.

[0048] S5, inputting the compressed feature data into a flexible AMOLED driving module, and adopting a time division multiplexing driving strategy to realize real-time display of images on the flexible AMOLED screen by loading image data of different regions at different times according to the compressed feature data.

[0049] Further, the multi-level feature extraction convolutional neural network module in S1 comprises at least three convolutional layers with different kernel sizes, wherein the first convolutional layer uses a large-size kernel for global feature extraction, and the subsequent convolutional layers use small-size kernels for local detailed feature extraction.

[0050] Further, the multi-level feature extraction convolutional neural network module further comprises a feature enhancement unit, which is composed of a spatial attention submodule and a channel attention submodule in cascade, wherein the spatial attention submodule first performs global average pooling and global maximum pooling operations on the feature map output by the convolutional layer, concatenates the two pooling results in the channel dimension, and inputs them into a multi-layer perceptron containing one hidden layer, to generate a spatial attention weight matrix through a Sigmoid activation function, and then performs weight adjustment in the spatial dimension on the feature map through element-wise multiplication; the channel attention submodule performs global average pooling and global maximum pooling in the width and height dimensions on the feature map respectively, inputs the pooling results into a shared multi-layer perceptron network, processes them through a ReLU activation function and a fully connected layer, and then generates a channel attention weight vector through a Sigmoid activation function, to perform weight scaling on the channel dimension of the feature map in a broadcast form; the weighted feature maps output by the two submodules are added element by element, and the resulting enhanced feature map is input into the next convolutional layer.

[0051] Further, the construction of the color mapping conversion convolutional neural network module in S3 is as follows:

[0052] The color response matrix is constructed by collecting the RGB sub-pixel brightness response data of the flexible AMOLED screen under different gray scales where m is the number of gray scales, and n is the number of measurements; the color conversion function is constructed according to the color response matrix, and the formula is: c (x) = W·x + b, where x is the pixel vector of the input feature map, is the conversion weight matrix obtained by least squares optimization according to the color response matrix R, is the bias vector; the color conversion function is embedded into the fully connected layer after the convolutional layer to perform color space conversion on the feature map set, wherein the pixel rearrangement process is realized by constructing an index matrix The pixels are re-indexed according to the arrangement mode of the flexible AMOLED screen to realize pixel rearrangement and obtain feature data suitable for display on the flexible AMOLED screen, where h and w are the height and width of the feature map, respectively.

[0053] Further, as Figure 2As shown, the color mapping conversion convolutional neural network module realizes efficient color space conversion and accurate pixel rearrangement of the feature map through a double-branch structure. The first branch adopts a U-Net structure with a hollow convolution to perform multi-scale feature fusion on the feature map, capturing local and global color information of the image. The second branch is a pixel arrangement prediction branch, which is composed of a convolution layer and an up-sampling layer, and predicts the offset Δp of the flexible AMOLED screen pixel arrangement ij , where i, j are pixel coordinates; the feature map output by the first branch is fused with the offset predicted by the second branch, and the formula is used to realize pixel rearrangement, where x r is the input feature map, y is the output feature data, and g is a bilinear interpolation function.

[0054] Further, the feature data adapted to the flexible AMOLED screen display is obtained in S3, which is obtained by inputting the feature map after color space conversion and pixel rearrangement processing into the adaptive gamut mapping layer. By calculating the color distribution histogram H(c) of the feature map and comparing it with the pre-stored flexible AMOLED screen gamut boundary B(c), the gamut mapping function is generated, where c represents the color channel; for color components that exceed the screen gamut, a gradient-preserving nonlinear compression algorithm C(c) = c·exp(-α·(c-B(c)) 2 is used for processing, where α is a compression coefficient dynamically adjusted according to the screen gamut coverage. The pixel rearranged feature map is geometrically corrected by the pixel arrangement compensation module to obtain the corrected feature map, which is then fused and reduced in dimension by a depth separable convolution layer to output the adapted feature data.

[0055] Further, the corrected feature map is input into a spatial attention-guided deep convolution layer through a depth separable convolution layer for feature fusion and dimension reduction. Spatial features are extracted for each channel, and a channel attention weight vector is generated through global average pooling. The feature map output by the deep convolution is multiplied by the attention weight vector channel by channel to obtain a spatial enhanced feature map F S . The channel number of the spatial enhanced feature map is reduced to the target channel number by a point convolution layer, which adopts a grouped convolution strategy to divide the input channels into target groups, and each group performs convolution independently to generate an intermediate feature map F r . Through a dynamic routing mechanism, the weight matrix W g (γ) of the grouped convolution is adaptively adjusted according to the gamut characteristic parameter γ of the flexible AMOLED screen, and the output adapted feature data is obtained, with the formula being: Optimal color representation is achieved on the RGB three channels, where represents the intermediate feature map of the lth group, and G is the target group number.

[0056] Further, the compression decision convolutional neural network module in S4 comprises a double-path structure, the main path is composed of cascaded residual convolutional blocks and gated recurrent units, the residual convolutional blocks extract local and global context information of the feature data through a hollow convolution, and the gated recurrent units perform time series modeling on the feature sequence to capture long-distance dependencies; the auxiliary path adopts a self-attention mechanism module to calculate an attention weight matrix of the feature data, and realizes multi-scale feature interaction through a multi-head attention mechanism; the output features of the two paths are weighted and summed through a feature fusion layer, the weight is dynamically generated by the display parameters of the flexible AMOLED screen, and the importance score vector of the output feature data is adaptively compressed.

[0057] Further, the dynamic threshold control strategy in S4 is as follows: the importance score vector of the feature data output by the compression decision convolutional neural network module is subjected to probability density estimation, the display resolution R and the color depth D of the flexible AMOLED screen are obtained, and the dynamic compression threshold T is calculated, the formula being: wherein β is a threshold adjustment factor, and the formula is: P(x) is a frequency distribution value, x is a score vector value, μ is a score vector mean, and σ is a standard deviation; the feature data with an importance score lower than the threshold T is quantitatively compressed, and the data higher than the threshold is losslessly compressed; and the threshold adjustment factor β is iteratively updated to realize dynamic balance between the compression rate and the image quality.

[0058] Further, the time-division multiplexing driving strategy of the flexible AMOLED driving module in S5 is as follows: the compressed feature data is divided into a plurality of non-overlapping region blocks O according to the spatial position, in each display period, the region block data is loaded in sequence according to the priority sequence, the high-priority region is loaded with full resolution, and the low-priority region is loaded with down-sampling; through time interleaving technology, the different region block data of adjacent display periods are time-division multiplexed, and the formula is: wherein D(t) is the driving data at time t, w k is a weight coefficient, p k(t) is the kth loaded region in the tth period, K is the number of regions loaded in each period, and the real-time display of the image on the flexible AMOLED screen is realized.

[0059] The embodiment describes in detail the flexible AMOLED screen, which is accurately adapted to the screen display through multi-level feature extraction and color mapping conversion, improves the image quality, restores the real color and clear details, balances the compression rate and the image quality by using the compression decision network and the dynamic threshold strategy, reduces the storage and transmission cost, realizes the real-time display of the image by using the time-division multiplexing driving strategy, optimizes the device performance, effectively solves the flexible AMOLED image display problem, improves the user visual experience, and promotes the development of related devices.

[0060] Based on embodiment 1, this embodiment describes in detail that the compression decision convolutional neural network module in S4 contains a double-path structure, as shown in the specific implementation of the double-path structure is as follows: Figure 3

[0061] The compression decision convolutional neural network module adopts a double-path structure, and the residual convolutional block and the gated recurrent unit in the main path are cascaded. The residual convolutional block uses the atrous convolution to capture the local and global context information of the feature data from different scales. The atrous convolution expands the receptive field of the convolution kernel without increasing the number of parameters and the amount of calculation, so that the network can obtain more extensive regional features and has stronger capture ability for complex information such as texture and shape in the image.

[0062] The gated recurrent unit focuses on the temporal modeling of the feature sequence to capture long-distance dependencies. When processing continuous feature data, it can remember important information at previous time points and dynamically adjust the memory content according to the current input, thereby effectively processing features with time correlation. When processing video image sequences, the gated recurrent unit can remember the key features of the previous frame and combine the information of the current frame for comprehensive analysis to accurately grasp the trend of image content changes. Through the processing of the main path, the network can deeply understand the characteristics of the feature data in the spatial and temporal dimensions, providing strong support for the subsequent importance evaluation.

[0063] The auxiliary path adopts a self-attention mechanism module, and the multi-head attention mechanism in it is the core of realizing multi-scale feature interaction. The multi-head attention mechanism uses multiple different attention heads to calculate attention weights in parallel from different representation subspaces, so as to simultaneously focus on different aspects of the feature data. Each attention head can focus on features of different scales, some focusing on local details and some focusing on global structures, thereby realizing comprehensive capture and interaction of multi-scale features.

[0064] For example, when processing a complex image containing multiple objects, different attention heads can focus on the features of different objects, some focusing on the outlines of the objects and some focusing on the textures of the objects, and then these multi-scale feature information is fused. The output features of the two paths are weighted and summed in the feature fusion layer, and the weights are dynamically generated by the display parameters of the flexible AMOLED screen. The display parameters include the resolution, refresh rate, color gamut, etc. of the screen, which will affect the final display effect of the image. According to the parameters, the fusion weights are dynamically adjusted, so that the importance score vector of the feature data output by the network is more in line with the actual display requirements, thereby realizing more accurate adaptive compression, improving the compression efficiency while ensuring the image quality.

[0065] ​The embodiment details a dual-path structure compression decision convolutional neural network module, the main path captures spatio-temporal features, the auxiliary path realizes multi-scale feature interaction, and then dynamically fuses according to the screen display parameters, accurately evaluates the importance of features, realizes adaptive compression, effectively balances the compression rate and image quality, reduces the data processing cost, and improves the overall performance of the flexible AMOLED screen image display.

[0066] Based on embodiment 1, the technical effects of the image compression rate, peak signal-to-noise ratio and structural similarity index of the technical scheme of the present application are described in detail in this embodiment, specifically:

[0067] By comparing the image compression rates of the present application and the prior art for different types of images such as natural scenery and people, the prior art selects the current mainstream flexible AMOLED image compression driving technology, uses a fixed color conversion matrix without optimization for the characteristics of the flexible AMOLED screen; the compression algorithm uses a general JPEG-like algorithm, which lacks a deep understanding of image features; the driving strategy is fixed frame rate line-by-line scanning, which does not take into account the differences in image content, and processes the same data set, records the compression rate, display quality and display efficiency data, and from Figure 4 It can be clearly seen from the above table that, taking natural scenery images as an example, the compression rate of the present application is 82.34%, while that of the prior art is only 69.17%, which means that when storing or transmitting natural scenery images, the present application can reduce the data volume by nearly two-thirds, while the prior art can only reduce less than half. For human images, the compression rate of the present application is 72.56%, and that of the prior art is 53.89%; in summary, the average compression rate of the present application is 77.42%, and that of the prior art is 61.53%, which is about 16 percentage points higher than that of the prior art.

[0068] The multi-level feature extraction convolutional neural network module of the present application can accurately extract the global and local features of the image, so that important and secondary information can be better distinguished in the subsequent compression process. The dynamic threshold control strategy adopted by the compression decision convolutional neural network module can dynamically adjust the compression threshold according to the frequency distribution of the feature data, improve the compression rate under the premise of ensuring the image quality, while the prior art lacks in-depth analysis of image features and targeted optimization for the characteristics of the flexible AMOLED screen, and cannot effectively reduce the data volume like the present application, resulting in a lower compression rate. Higher compression rate not only saves a lot of storage space, but also reduces transmission time and bandwidth requirements during data transmission, which is of great significance to improving the overall performance of the device.

[0069] Meanwhile, in terms of PSNR, the test results of natural landscape images for the method of the application are 38.56dB, and the prior art is only 32.45dB, with a difference of 6.11dB. The higher the PSNR value, the smaller the distortion of the image, indicating that when processing natural landscape images, the method of the application can better preserve the details and color information of the image and reduce the loss of image quality caused by compression and driving process; for portrait images, the PSNR of the method of the application is 37.89dB, and the prior art is 31.78dB; as Figure 5 As shown in the peak signal-to-noise ratio comparison chart, the average PSNR of the method of the application is 38.22dB, and the prior art is 32.11dB.

[0070] Structural similarity index (SSIM), which measures the similarity of images from three aspects of brightness, contrast and structure, has a value range of 0 to 1, and the closer to 1, the more similar the image and the higher the quality. In the natural landscape image test, the SSIM of the method of the application is 0.964, while the prior art is only 0.902, with a significant gap. In terms of portrait images, the SSIM of the method of the application is 0.957, and the prior art is 0.896; as Figure 6 As shown in the peak signal-to-noise ratio comparison chart, the average PSNR of the method of the application is 38.22dB, and the prior art is 32.11dB.

[0071] The excellent performance of the method of the application in PSNR and SSIM is due to the careful design of the color mapping conversion convolutional neural network module. By collecting the pixel arrangement mode and color response characteristic data of the flexible AMOLED screen, a precise color conversion function and pixel rearrangement mechanism are constructed, which can better restore the true color and details of the image. In the compression process, the dynamic threshold control strategy avoids the decline of image quality caused by excessive compression; in contrast, the prior art lacks deep adaptation to the characteristics of the screen and effective image quality protection mechanism, and is prone to color deviation, detail loss and other problems when processing images, resulting in lower PSNR and SSIM. These data fully demonstrate that the method of the application can provide higher quality image display effect for the flexible AMOLED screen, and bring users a clearer and more realistic visual experience.

[0072] This embodiment describes in detail that the application technology has remarkable achievements in flexible AMOLED image compression and driving. Through the unique design of the convolutional neural network module, efficient feature extraction, precise color conversion and intelligent compression decision are realized. Compared with the prior art, the compression rate is improved by about 17%, the display quality is greatly improved, and the display performance is optimized in all aspects.

[0073] The above merely describes the preferred embodiments of the present application, and is not intended to limit the protection scope of the present application. The present application can have various changes and modifications for those skilled in the art; any change, modification, replacement, integration and parameter change of the embodiments within the spirit and principle of the present application, by conventional substitution or capable of realizing the same function without departing from the principle and spirit of the present application, all fall within the protection scope of the present application.

Claims

1. A flexible AMOLED image compression driving method based on convolutional neural network, characterized in that: The following steps are involved: S1. Obtaining original image data to be displayed, and inputting the original image data into a pre-trained multi-level feature extraction convolutional neural network module; S2. Extract features from the original image data using the multi-level feature extraction convolutional neural network module to obtain a feature map set containing global features and local detail features; S3. Constructing a color mapping conversion convolutional neural network module based on the pixel arrangement and color response characteristics of the flexible AMOLED screen, inputting the feature map set into the color mapping conversion convolutional neural network module, and performing color space conversion and pixel rearrangement processing on the feature map set by the color mapping conversion convolutional neural network module to obtain feature data adapted for display on the flexible AMOLED screen; S4. Constructing a compression decision convolutional neural network module, wherein the compression decision convolutional neural network module adaptively compresses the feature data according to the feature importance distribution of the feature data adapted for display on the flexible AMOLED screen. During the compression process, a dynamic threshold control strategy is adopted to dynamically adjust the compression threshold according to the frequency distribution of the feature data to balance the compression rate and image quality. S5. Input the compressed feature data into the flexible AMOLED driving module. Based on the compressed feature data, a time-sharing multiplexing driving strategy is adopted to load image data of different areas in a time-sharing manner to achieve real-time display of the image on the flexible AMOLED screen.

2. The flexible AMOLED image compression driving method based on convolutional neural network according to claim 1, characterized in that: The multi-level feature extraction convolutional neural network module in S1 includes at least three convolution layers with different convolution kernel sizes, wherein the first convolution layer uses a large-size convolution kernel for global feature extraction, and the subsequent convolution layers use small-size convolution kernels for local detail feature extraction.

3. The flexible AMOLED image compression driving method based on convolutional neural network according to claim 2, characterized in that: The multi-level feature extraction convolutional neural network module also includes a feature enhancement unit, which is composed of a cascade of a spatial attention submodule and a channel attention submodule, wherein the spatial attention submodule first performs global average pooling and global maximum pooling operations on the feature map output by the convolution layer, and inputs the two pooling results into a multi-layer perceptron containing a hidden layer after splicing them in the channel dimension, and generates a spatial attention weight matrix through a Sigmoid activation function, and adjusts the weight of the spatial dimension of the feature map by element multiplication; the channel attention submodule performs global average pooling and global maximum pooling on the feature map in the width and height dimensions respectively, and inputs the pooling results into a shared multi-layer perceptron network, and after processing by a ReLU activation function and a fully connected layer, generates a channel attention weight vector through a Sigmoid activation function, and performs weight scaling on the channel dimension of the feature map in a broadcast form; The weighted feature maps output by the two sub-modules are added element-by-element and input into the next convolutional layer as the enhanced feature maps.

4. The flexible AMOLED image compression driving method based on convolutional neural network according to claim 1, characterized in that: The construction of the color mapping conversion convolutional neural network module in S3 is specifically as follows: By collecting the RGB sub-pixel brightness response data of the flexible AMOLED screen at different grayscales, a color response matrix is ​​constructed. Where m is the number of gray levels and n is the number of measurements; based on the color response matrix, a color conversion function is constructed, and the formula is: c (x) = W x + b, where x is the pixel vector of the input feature map, is the conversion weight matrix obtained by least squares optimization based on the color response matrix R, is the bias vector; the color conversion function is embedded into the fully connected layer after the convolution layer, and the color space conversion is performed on the feature map set, wherein the pixel rearrangement process is performed by constructing the index matrix The pixels are reindexed according to the arrangement of the flexible AMOLED screen to achieve pixel rearrangement and obtain feature data adapted to the display of the flexible AMOLED screen, where h and w are the height and width of the feature map respectively.

5. The flexible AMOLED image compression driving method based on convolutional neural network according to claim 4, characterized in that: The color mapping conversion convolutional neural network module achieves efficient color space conversion and precise pixel rearrangement of feature maps through a dual-branch structure. The first branch uses a U-Net structure with void convolution to perform multi-scale feature fusion on the feature map to capture the local and global color information of the image. The second branch is the pixel arrangement prediction branch, which predicts the offset Δp of the flexible AMOLED screen pixel arrangement through a structure composed of a convolution layer and an upsampling layer. ij , where itj is the pixel coordinate; the feature map output by the first branch is fused with the offset predicted by the second branch, and the formula To achieve pixel rearrangement, where x r is the input feature map, y is the output feature data, and g is the bilinear interpolation function.

6. The flexible AMOLED image compression driving method based on convolutional neural network according to claim 5, characterized in that: In S3, feature data adapted to the display of the flexible AMOLED screen is obtained, and the feature map after color space conversion and pixel rearrangement is input into the adaptive color gamut mapping layer. The color distribution histogram H(c) of the feature map is calculated and compared with the pre-stored flexible AMOLED screen color gamut boundary B(c) to generate a color gamut mapping function. Where c represents the color channel; for color components that exceed the screen color gamut, a gradient-preserving nonlinear compression algorithm C(c) = c exp(-α (cB(c)) is used. 2 ) for processing, where α is the compression coefficient dynamically adjusted according to the screen color gamut coverage. The feature map after pixel rearrangement is geometrically corrected by the pixel arrangement compensation module to obtain the corrected feature map. The feature fusion and dimensionality reduction are performed through the depthwise separable convolution layer to output the adapted feature data.

7. The flexible AMOLED image compression driving method based on convolutional neural network according to claim 6, characterized in that: The feature fusion and dimensionality reduction are performed through the depth-wise separable convolution layer, and the corrected feature map is input into the depth-wise convolution layer guided by spatial attention, and spatial features are extracted for each channel. The channel attention weight vector is generated by global average pooling, and the feature map output by the depth-wise convolution is multiplied by the attention weight vector channel by channel to obtain the spatial enhanced feature map F. S , the number of channels of the spatial enhancement feature map is reduced to the target number of channels through the point convolution layer. The point convolution layer adopts the group convolution strategy to divide the input channels into target groups, and each group is convolved independently to generate the intermediate feature map F r , through the dynamic routing mechanism, according to the color gamut characteristic parameter γ of the flexible AMOLED screen, the weight matrix W of the group convolution is adaptively adjusted g (γ), the output adaptation feature data, the formula is: Achieve optimal color representation on the three RGB channels, where represents the intermediate feature map of the lth group, and G is the number of target groups.

8. The flexible AMOLED image compression driving method based on convolutional neural network according to claim 1, characterized in that: The compression decision convolutional neural network module in the S4 includes a dual-path structure. The main path consists of cascaded residual convolution blocks and gated recurrent units. The residual convolution blocks extract local and global context information of feature data through void convolution, and the gated recurrent units perform temporal modeling on feature sequences to capture long-distance dependencies. The auxiliary path adopts a self-attention mechanism module to calculate the attention weight matrix of feature data and realize multi-scale feature interaction through a multi-head attention mechanism. The output features of the two paths are weighted and summed through a feature fusion layer. The weights are dynamically generated by the display parameters of the flexible AMOLED screen, and the importance score vector of the output feature data is adaptively compressed.

9. The flexible AMOLED image compression driving method based on convolutional neural network according to claim 1 or 8, characterized in that: The dynamic threshold control strategy in S4 is specifically as follows: performing probability density estimation on the feature data importance score vector output by the compression decision convolutional neural network module, obtaining the display resolution R and color depth D of the flexible AMOLED screen, and calculating the dynamic compression threshold T. The formula is: Where β is the threshold adjustment factor, the formula is: P(x) is the frequency distribution value, x is the score vector value, μ is the score vector mean, and σ is the standard deviation. Feature data with importance scores lower than the threshold T are quantized and compressed, and data above the threshold are losslessly compressed. The threshold adjustment factor β is iteratively updated to achieve a dynamic balance between compression rate and image quality.

10. The flexible AMOLED image compression driving method based on convolutional neural network according to claim 1, characterized in that: The time-sharing multiplexing driving strategy of the flexible AMOLED driving module in S5 is as follows: the compressed feature data is divided into multiple non-overlapping area blocks O according to spatial position. In each display cycle, the area block data is loaded in sequence according to the priority sequence. The high-priority area is loaded with full resolution, and the low-priority area is loaded with downsampling. The different area block data of adjacent display cycles are time-domain multiplexed through time interleaving technology. The formula is: Where D(t) is the driving data at time t, w k is the weight coefficient, p k(t) is the kth loaded area in the tth cycle, K is the number of areas loaded in each cycle, and the real-time display of the image on the flexible AMOLED screen is realized.

Images