Image processing method, system, electronic device and computer readable storage medium

Abstract

The application discloses an image processing method, a system, an electronic device and a computer readable storage medium. The method comprises the following steps: dividing an original image into a JND point image and a residual image; inputting the JND point image and the residual image into a preset multi-level residual compensation network model, and outputting to obtain latent representation parameters; and sequentially quantizing and entropy encoding the latent representation parameters to obtain a target compressed image corresponding to the original image. The application improves the compression effect of image compression.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an image processing method, system, electronic device, and computer-readable storage medium. Background Technology

[0002] When performing image compression, the traditional HEVC-SCC (High Efficiency Video Coding-Screen Content Coding) method relies on manually created screen content features and traditional optimized rates to minimize rate distortion loss. However, because it is a manual operation, the compression efficiency is severely limited, resulting in poor compression performance. Summary of the Invention

[0003] The main objective of this invention is to provide an image processing method, system, electronic device, and computer-readable storage medium, aiming to solve the technical problem of how to improve the compression effect of image compression.

[0004] To achieve the above objectives, this application provides an image processing method, comprising:

[0005] The original image is segmented into a JND point image and a residual image;

[0006] The JND point image and the residual image are input into a preset multi-level residual compensation network model, and the latent representation parameters are output.

[0007] The latent representation parameters are sequentially quantized and entropy encoded to obtain the target compressed image corresponding to the original image.

[0008] Optionally, the multi-level residual compensation network model includes two sub-encoder models sharing the same network structure, and the network model formulas corresponding to the moduli of the two sub-encoders respectively include:

[0009]

[0010] Wherein, the x m Represents the residual image, where x p This represents the JND point image, where i represents the number of processing steps. This represents the residual image from the previous processing stage that is processed by the multi-level residual compensation network model. The image represents the JND point image processed in the previous step of the multi-level residual compensation network model, where ⊙ represents element-wise multiplication.

[0011] Optionally, the step of inputting the JND point image and the residual image into a preset multi-level residual compensation network model and outputting the latent representation parameters includes:

[0012] The residual image is input into a sub-encoder model for forward transformation processing to obtain the first transformation processing result;

[0013] The JND point image and the first transformation result are subjected to a forward transformation process by another sub-encoder to obtain the second transformation result.

[0014] The potential representation parameters are determined based on the preset number of processing steps and the result of the second transformation process.

[0015] Optionally, the step of segmenting the original image into a JND point image and a residual image includes:

[0016] Identify the distorted image in the JND dataset that corresponds to the original image;

[0017] Determine the JND point image based on the original image and the distorted image;

[0018] The residual image is determined based on the JND point image and the original image.

[0019] Optionally, before the step of determining the distorted image in the JND dataset corresponding to the original image, the following steps are included:

[0020] Construct a data table containing the correspondence between original images and distorted images, and use the data table as the JND dataset. The data table includes the correspondence between at least one original image and at least one distorted image, wherein the distorted image is obtained by compressing the original image.

[0021] Optionally, the step of determining the JND point image based on the original image and the distorted image includes:

[0022] The distorted images are compared sequentially with the original image;

[0023] If among the multiple distorted images there exists a distorted image that has a visual difference from the original image, then the distorted image with the visual difference is taken as the JND point image, wherein the visual difference includes image difference pixels that can be clearly identified by the human eye.

[0024] Optionally, determining the residual image based on the JND point image and the original image includes:

[0025] The pixel values ​​at the same locations between the original image and the JND point image are subtracted, and the original image after subtracting the pixel values ​​at all locations is used as the residual image.

[0026] In addition, to achieve the above objectives, this application also provides an image processing system, which includes a preprocessing module, an analysis and transformation module, and a quantization and encoding module;

[0027] The preprocessing module is used to segment the original image into a JND point image and a residual image;

[0028] The analysis and transformation module is used to input the JND point image and the residual image into a preset multi-level residual compensation network model and output the potential representation parameters.

[0029] The quantization and encoding module is used to sequentially quantize and entropy encode the latent representation parameters to obtain the target compressed image corresponding to the original image.

[0030] In addition, to achieve the above objectives, this application also provides an electronic device, which includes: a memory, a processor, and an image processing program stored in the memory and executable on the processor. When the image processing program is executed by the processor, it implements the steps of the image processing method described above.

[0031] In addition, to achieve the above objectives, this application also provides a computer-readable storage medium storing an image processing program, which, when executed by a processor, implements the steps of the image processing method described above.

[0032] This application embodiment segments the original image into JND point images and residual images using Justify distortion guidance, without relying on manual feature extraction. This makes the segmented JND point images as close as possible to the original image. During image compression, the JND point images and residual images are input together into a multi-level residual compensation network model, and the output is the latent representation parameters. Then, the latent representation parameters are quantized and entropy encoded to obtain the target compressed image. In other words, by combining the JND point images and residual images for transformation processing, the bits can be adaptively allocated when generating the target compressed image based on the residual image, thereby improving the compression effect of the image compression. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the first embodiment of the image processing method of this application;

[0034] Figure 2 This is a flowchart illustrating the second embodiment of the image processing method of this application;

[0035] Figure 3 This is a schematic diagram of the module framework of the image processing system of this application;

[0036] Figure 4 This is a schematic diagram of the overall process of the image processing method in this application;

[0037] Figure 5 This is a schematic diagram of the preprocessing module in the image processing method of this application;

[0038] Figure 6 This is a schematic diagram of the multi-level residual compensation network model in the image processing method of this application;

[0039] Figure 7 This is a schematic diagram of the device structure of the hardware operating environment involved in the image processing method in the embodiments of this application.

[0040] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0041] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0042] Because HEVC-SCC (High Efficiency Video Coding-Screen Content Coding) suffers from low compression efficiency and poor image compression performance, this embodiment employs just-observable distortion-guided end-to-end screen content image compression. It also utilizes prior information about human visual perception to adaptively allocate bits, thus avoiding poor compression performance for screen content images. Furthermore, the reconstructed image using just-observable distortion-guided end-to-end screen content image compression better conforms to human visual perception patterns, and the ability to adaptively allocate bits using prior information about human visual perception significantly improves compression performance.

[0043] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0044] Reference Figure 1 This application provides an image processing method. In a first embodiment of the image processing method, the image processing method is applied to an image processing system and includes:

[0045] Step S10: Segment the original image into a JND point image and a residual image;

[0046] Step S20: Input the JND point image and the residual image into a preset multi-level residual compensation network model, and output the potential representation parameters.

[0047] Step S30: The latent representation parameters are sequentially quantized and entropy encoded to obtain the target compressed image corresponding to the original image.

[0048] In this embodiment, image compression can be achieved through image encoding. When compressing a frame of image, the frame is first divided into multiple image blocks. Each image block, or a frame of image, is subtracted from the predicted block obtained through a prediction mode to obtain a residual block. This residual block is then transformed and quantized, and encoded using an entropy encoder to form an encoded bitstream, which is then sent to the decoder. The decoder reconstructs the image using a synthetic transform after decoding. Optionally, the image compression or image encoding / decoding process can include a context model of analytical transform, synthetic transform, quantization, and entropy encoding. This embodiment adds a preprocessing module based on Just-Observable Distortion (JND-ESN) and a multi-level residual compensation structure. Therefore, the original input image x can be fed into the preprocessing module to eliminate human visual redundancy. Then, the output of the preprocessing module is converted into a latent representation y through the analytical transform module, which includes a multi-level residual compensation structure. Additive noise is then used to approximate the latent representation y instead of quantization to obtain compressed data. Entropy encoding is then used to encode the compressed data to obtain a bitstream, which is then sent to the decoder. After decoding the bitstream at the decoding end, a synthesis transform is used to reconstruct the image from the decoded data. For example, ... Figure 4 As shown, the original image is input into the JND-ESN, and the output is a residual image and a JND point image. These images are then input into the analysis and transformation module, where they are processed by a multi-level residual compensation network model to obtain latent representation parameters. These parameters are then processed through two encoding branches. The first branch processes the data through a quantization module and an entropy coding module to obtain a bitstream, which is then decoded by an entropy decoder at the decoding end, and entropy estimation is performed to obtain the entropy estimation result. The second branch processes the data through a quantization module and an entropy coding module, and the processing result is encoded together with the entropy estimation result to obtain another bitstream. This bitstream is then decoded by an entropy decoder at the decoding end to obtain the reconstructed image.

[0049] Further, in step S10, the original image is segmented into a JND point image and a residual image;

[0050] In this embodiment, when image compression is required, the original image to be compressed is first acquired, and then the original image is segmented to obtain a JND point image and a residual image. The original image can be a single frame or a segmented image block; there are no restrictions. The JND point image is the image generated after processing the original image using a pre-set JND model. The residual image is the JND image generated by the original processing module after processing the original image; this JND image contains the energy removed by the pre-processing module.

[0051] The Just Noticeable Distortion (JND) is used to represent the maximum image distortion that the human eye cannot perceive, reflecting the human eye's tolerance for image changes.

[0052] In this embodiment, the original image can be segmented using a preprocessing module into a JND point image and a residual image. Optionally, the preprocessing module can be configured as JND-ESN, which generates a low-energy image by reducing perceptual redundancy, i.e., an image after removing imperceptible signal components, such as... Figure 5 The diagram shows the JND-ESN framework for the preprocessing module, which includes the GT dataset, the original image, the preprocessing module, JND point images, and residual images. Furthermore, the preprocessing module can make the image as close as possible to the original image, thus enabling JND-ESN to achieve perceptually lossless image preprocessing.

[0053] It's important to note that the residual image includes the energy removed by the preprocessing module. A higher pixel value at a given location in the residual image indicates a higher level of distortion at that location. Therefore, in subsequent processing, these regions experience greater compression distortion at the same bitrate, requiring compensation. This can be achieved using a multi-level residual compensation network model to compensate for pixel values.

[0054] For example, if there exists an original image x ori If compression is required, the original image x can be compressed using a preprocessing module. ori Segmented into predicted JND point image x p and residual image x m Among them, the JND point image x p and residual image x m They are represented as follows:

[0055] x p =E(x) ori );

[0056] x m =x ori -x p;

[0057] Where E() represents the ESN-JND network processing procedure. E(x) ori This indicates that the original image was processed by the ESN-JND network.

[0058] Further, in step S20, the JND point image and the residual image are input into a preset multi-level residual compensation network model, and the potential representation parameters are output.

[0059] In this embodiment, after obtaining the JND point image and residual image through the preprocessing module, the JND point image and residual image can be combined for analysis and transformation processing to obtain latent representation parameters. The latent representation parameters can be a compressed hidden vector transformed from the image by the encoder. The analysis and transformation process requires network processing of the JND point image and residual image using a multi-level residual compensation network model.

[0060] Furthermore, the multi-level residual compensation network model includes two sub-encoder models sharing the same network structure, and the network model formulas corresponding to the moduli of the two sub-encoders include:

[0061]

[0062] Wherein, the x m Represents the residual image, where x p This represents the JND point image, where i represents the number of processing steps. This represents the residual image from the previous processing stage that is processed by the multi-level residual compensation network model. The image represents the JND point image processed in the previous step of the multi-level residual compensation network model, where ⊙ represents element-wise multiplication.

[0063] For example, such as Figure 6 As shown, the multi-level residual compensation network model includes two sub-encoder models sharing the same network structure. The sub-encoder models can have multi-layered network architectures, and the number of layers is related to 'i' in the network model formula; that is, 'i' in the network model formula can be set to equal the number of layers in the network architecture. For example... Figure 6As shown, a four-layer network architecture is used, where i can be 4. The first layer is RBS, the second layer is RB and RBS, the third layer is CBAM and RBS, and the fourth layer is RB and RBS. RB includes Conv (convolutional layer), GDN (normalization layer), LRELU (activation function layer), Conv, GDN, and LRELU. RBS includes Conv With Stride, GDN, LRELU, Conv With Stride, GDN, and LRELU. Furthermore, RB represents a residual block, and RBS represents a downsampling operation based on RB.

[0064] In this embodiment, when multiple transformations are required for the residual image and the JND point image, the pixels in the residual image can be transformed once to obtain a transformation result. When transforming the same pixel in the JND point image, the transformation result of the first transformation and the transformation result of the same pixel in the previous transformation are combined to perform the transformation on the same pixel in the JND point image. This achieves compensation for areas with high distortion levels in the JND point image and achieves the purpose of adaptive bit allocation based on the residual image.

[0065] Further, step S20, which involves inputting the JND point image and the residual image into a preset multi-level residual compensation network model and outputting the latent representation parameters, includes:

[0066] Step a: Input the residual image into a sub-encoder model for forward transform processing to obtain the first transform processing result;

[0067] Step b: Perform a forward transformation on the JND point image and the first transformation result using another sub-encoder to obtain the second transformation result;

[0068] Step c: Determine the potential representation parameters based on the preset number of processing times and the result of the second transformation process.

[0069] In this embodiment, when performing analysis and transformation processing on the JND point image and residual image, the number of analysis and transformation processes required can be determined first, and then the JND point image and residual image are subjected to the analysis and transformation processing according to the required number of processes. During the analysis and transformation processing, the residual image can be input into a sub-encoder for the first forward transformation process to obtain the first residual transformation result. The JND point image is then input into another sub-encoder for the first forward transformation process to obtain the first JND transformation result. During the second transformation process, the first residual transformation result is directly subjected to the second forward transformation process through a sub-encoder until the required number of processes is reached. The network model formula corresponding to one sub-encoder can be...

[0070] When performing a second transformation on the first JND transform result, the first residual transform result and the first JND transform result need to be multiplied element-wise. Then, another sub-encoder combines the result of the element-wise multiplication with the first JND transform result for a second forward transform, and this process is repeated until the required number of transformations is reached. The network model formula for the other sub-encoder can be:

[0071]

[0072] Furthermore, the result of the final transformation process of the two sub-encoders can be used as a potential representation parameter.

[0073] Further, in step S30, the latent representation parameters are sequentially quantized and entropy encoded to obtain the target compressed image corresponding to the original image.

[0074] In this embodiment, after obtaining the latent representation parameters, the latent representation parameters can be quantized, such as by using additive noise to approximate the quantization, and then entropy coding is performed after quantization to achieve the target compressed image corresponding to the original image. In other words, the compression of the original image is now complete.

[0075] In this embodiment, the original image is segmented into JND point images and residual images by Justifiable distortion guidance, without relying on manual feature extraction. This makes the segmented JND point images as close as possible to the original image. During image compression, the JND point images and residual images are input together into a multi-level residual compensation network model, and the output is the latent representation parameters. Then, the latent representation parameters are quantized and entropy encoded to obtain the target compressed image. In other words, by combining the JND point images and residual images for transformation processing, the bits can be adaptively allocated when generating the target compressed image based on the residual image, thereby improving the compression effect of the image compression.

[0076] Furthermore, based on the first embodiment described above, a second embodiment of the image processing method of this application is proposed, referring to... Figure 2 In the second embodiment, step S10, the step of segmenting the original image into a JND point image and a residual image, includes:

[0077] Step d: Determine the distorted image in the JND dataset that corresponds to the original image;

[0078] Step e: Determine the JND point image based on the original image and the distorted image;

[0079] Step f: Determine the residual image based on the JND point image and the original image.

[0080] In this embodiment, when segmenting the original image into JND point images and residual images, all distorted images in the JND dataset that match the original image are first identified. Then, the original image and all distorted images are input into a preprocessing module. This module compares each distorted image with the original image to select a suitable distorted image as the JND point image. The residual image is then determined by comparing the original image with the JND point images. Finally, the original image can be obtained by merging the residual image and the JND point image.

[0081] In this embodiment, by determining the distorted image in the JND dataset that corresponds to the original image, determining the JND point image based on the original image and the distorted image, and then determining the residual image based on the JND point image and the original image, the original image is segmented into the JND point image and the residual image.

[0082] Further, before step d, which involves determining the distorted image in the JND dataset corresponding to the original image, the following steps are included:

[0083] Step x: Construct a data table containing the correspondence between original images and distorted images, and use the data table as the JND dataset. The data table includes the correspondence between at least one original image and at least one distorted image, wherein the distorted image is obtained by compressing the original image.

[0084] In this embodiment, before segmenting the original image into JND point images and residual images by the preprocessing module, it is necessary to construct a JND dataset corresponding to the original images. Multiple original images can be acquired in advance, and a compressed image corresponding to each original image can be obtained, which is then used as the distortion image. The degree of distortion in each distortion image corresponding to the original image is different. A blank data table can then be constructed, filled with all the original images, and then sequentially filled with the corresponding original and distortion images according to their correspondence, resulting in a data table showing the correspondence between original and distortion images. This data table is then used as the JND dataset.

[0085] In this embodiment, by constructing a data table containing the correspondence between the original image and the distorted image, and using the data table as the JND dataset, it is easier to determine the JND point image based on the JND dataset in the future.

[0086] Further, step e, the step of determining the JND point image based on the original image and the distorted image, includes:

[0087] Step e1: Compare the multiple distorted images with the original image in sequence;

[0088] Step e2: If there is a distorted image among the multiple distorted images that has a visual difference from the original image, then the distorted image with the visual difference is taken as the JND point image, wherein the visual difference includes image difference pixels that can be clearly identified by the user's human eye.

[0089] In this embodiment, multiple distorted images can be compared sequentially with the original image to determine whether there is a significant visual difference between the distorted image and the original image. When a distorted image with a significant visual difference from the original image is identified, this distorted image can be directly used as the JND point image. The method for determining whether there is a distorted image with a visual difference from the original image among multiple distorted images can include at least one of the following:

[0090] Method 1: Place each distorted image and the original image simultaneously on the screen for comparison. Upon detecting a user-inputted matching or mismatch command, replace the distorted image on the screen until all distorted images have been compared. Count all distorted images corresponding to the mismatch command, and select the distorted image that is closest to the original image from all distorted images corresponding to the mismatch command as the JND point image.

[0091] Method 2: Determine the degree of distortion of each distorted image relative to the original image, and take the distorted image corresponding to the degree of distortion within the preset threshold range as the target distorted image, and select one from the target distorted images as the JND point image.

[0092] Method 3: Place the distorted images on the screen in sequence according to the degree of distortion of each distorted image relative to the original image, so as to compare them with the original image on the screen. After detecting the user's input of the comparison inconsistency command, use the distorted image displayed on the screen when the user inputs the comparison inconsistency command as the JND point image.

[0093] In this embodiment, by comparing multiple distorted images sequentially with the original image, and using the distorted images that have visual differences from the original image as JND point images, it can be ensured that the acquired JND point images fit the original image as closely as possible, while also being able to adaptively allocate bits during subsequent image compression based on human visual perception learning.

[0094] Further, step f, the step of determining the residual image based on the JND point image and the original image, includes:

[0095] Step f1: Subtract the pixel values ​​of the same locations between the original image and the JND point image, and use the original image after subtracting the pixel values ​​of all locations as the residual image.

[0096] In this embodiment, after determining the JND point image, the original image and the JND point image can be input into the preprocessing module for processing. For example, each pixel in the original image can be subtracted from the pixel in the JND point image, and the original image after subtracting the pixel values ​​at all locations is used as the residual image. It should be noted that in the residual image, the higher the pixel value, the higher the distortion level at that pixel location.

[0097] In this embodiment, the residual image is obtained by subtracting all pixels in the original image from all pixels in the JND point image, thereby ensuring the validity of the obtained residual image.

[0098] In addition, refer to Figure 3 This application also provides an image processing system, which includes a preprocessing module A10, an analysis and transformation module A20, and a quantization and encoding module A30.

[0099] The preprocessing module A10 is used to segment the original image into a JND point image and a residual image;

[0100] The analysis and transformation module A20 is used to input the JND point image and the residual image into a preset multi-level residual compensation network model and output the potential representation parameters.

[0101] The quantization encoding module A30 is used to quantize and entropy encode the latent representation parameters sequentially to obtain the target compressed image corresponding to the original image.

[0102] Optionally, the multi-level residual compensation network model includes two sub-encoder models sharing the same network structure, and the network model formulas corresponding to the moduli of the two sub-encoders respectively include:

[0103]

[0104]

[0105] Wherein, the x m Represents the residual image, where x p This represents the JND point image, where i represents the number of processing steps. This represents the residual image from the previous processing stage that is processed by the multi-level residual compensation network model. The image represents the JND point image processed in the previous step of the multi-level residual compensation network model, where ⊙ represents element-wise multiplication.

[0106] Optionally, the analysis and transformation module A20 is used for:

[0107] The residual image is input into a sub-encoder model for forward transformation processing to obtain the first transformation processing result;

[0108] The JND point image and the first transformation result are subjected to a forward transformation process by another sub-encoder to obtain the second transformation result.

[0109] The potential representation parameters are determined based on the preset number of processing steps and the result of the second transformation process.

[0110] Optionally, the preprocessing module A10 is used for:

[0111] Identify the distorted image in the JND dataset that corresponds to the original image;

[0112] Determine the JND point image based on the original image and the distorted image;

[0113] The residual image is determined based on the JND point image and the original image.

[0114] Optionally, the preprocessing module A10 is used for:

[0115] Construct a data table containing the correspondence between original images and distorted images, and use the data table as the JND dataset. The data table includes the correspondence between at least one original image and at least one distorted image, wherein the distorted image is obtained by compressing the original image.

[0116] Optionally, the preprocessing module A10 is used for:

[0117] The distorted images are compared sequentially with the original image;

[0118] If among the multiple distorted images there exists a distorted image that has a visual difference from the original image, then the distorted image with the visual difference is taken as the JND point image, wherein the visual difference includes image difference pixels that can be clearly identified by the human eye.

[0119] Optionally, the preprocessing module A10 is used for:

[0120] The pixel values ​​at the same locations between the original image and the JND point image are subtracted, and the original image after subtracting the pixel values ​​at all locations is used as the residual image.

[0121] In addition, this application also provides an electronic device, which includes a memory, a processor, and an image processing program stored in the memory and executable on the processor. When the image processing program is executed by the processor, it implements the steps of the image processing method described above.

[0122] Furthermore, in one embodiment, Figure 7 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention, as shown below. Figure 7As shown, at the hardware level, this electronic device includes a processor, and optionally also includes an internal bus, a network interface, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk storage device. Of course, this electronic device may also include other hardware required for its functions. The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, Figure 7 The symbol is represented by only one double-headed arrow, but this does not imply that there is only one bus or one type of bus. Memory is used to store programs. Specifically, the program may include program code, which includes computer operation instructions. The processor reads the corresponding computer program from non-volatile memory into memory and then runs it, forming a shared resource access control device at the logical level. The processor executes the program stored in memory and specifically performs the steps of the aforementioned image processing method.

[0123] The specific implementation of the electronic device in this application is basically the same as the various embodiments of the image processing method described above, and will not be repeated here.

[0124] In addition, to achieve the above objectives, this application also provides a computer-readable storage medium storing an image processing program, which, when executed by a processor, implements the steps of the image processing method described above.

[0125] The specific implementation of the computer-readable storage medium in this application is basically the same as the various embodiments of the image processing method described above, and will not be repeated here.

[0126] It will be understood by those skilled in the art that all or some of the steps, systems, or apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0127] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. An image processing method, characterized in that, The image processing method includes: The original image is segmented into a Just-Observable Distortion (JND) point image and a residual image; The JND point image and the residual image are input into a preset multi-level residual compensation network model, and the latent representation parameters are output. The latent representation parameters are sequentially quantized and entropy encoded to obtain the target compressed image corresponding to the original image; The multi-level residual compensation network model includes two sub-encoder models sharing the same network structure. The network model formulas corresponding to the two sub-encoder models are as follows: Wherein, the x m Represents the residual image, where x p This represents the JND point image, where i represents the number of processing steps. This represents the residual image from the previous processing stage that is processed by the multi-level residual compensation network model. The image represents the JND point image processed in the previous step of the multi-level residual compensation network model, where ⊙ represents element-wise multiplication. The step of inputting the JND point image and the residual image into a preset multi-level residual compensation network model and outputting the latent representation parameters includes: The residual image is input into a sub-encoder model for forward transformation processing to obtain the first transformation processing result; The JND point image and the first transformation result are subjected to a forward transformation process by another sub-encoder to obtain the second transformation result. The potential representation parameters are determined based on the preset number of processing steps and the result of the second transformation process.

2. The image processing method as described in claim 1, characterized in that, The step of segmenting the original image into a JND point image and a residual image includes: Identify the distorted image in the JND dataset that corresponds to the original image; Determine the JND point image based on the original image and the distorted image; The residual image is determined based on the JND point image and the original image.

3. The image processing method as described in claim 2, characterized in that, Before the step of determining the distorted image in the JND dataset corresponding to the original image, the following steps are included: Construct a data table containing the correspondence between original images and distorted images, and use the data table as the JND dataset. The data table includes the correspondence between at least one original image and at least one distorted image, wherein the distorted image is obtained by compressing the original image.

4. The image processing method as described in claim 2, characterized in that, The step of determining the JND point image based on the original image and the distorted image includes: The distorted images are compared sequentially with the original image; If among the multiple distorted images there exists a distorted image that has a visual difference from the original image, then the distorted image with the visual difference is taken as the JND point image, wherein the visual difference includes image difference pixels that can be clearly identified by the human eye.

5. The image processing method as described in claim 2, characterized in that, The step of determining the residual image based on the JND point image and the original image includes: The pixel values ​​at the same locations between the original image and the JND point image are subtracted, and the original image after subtracting the pixel values ​​at all locations is used as the residual image.

6. An image processing system, characterized in that, The image processing system includes a preprocessing module, an analysis and transformation module, and a quantization and encoding module; The preprocessing module is used to segment the original image into a JND point image and a residual image; The analysis and transformation module is used to input the JND point image and the residual image into a preset multi-level residual compensation network model and output latent representation parameters; wherein, the multi-level residual compensation network model includes two sub-encoder models sharing the same network structure, and the network model formulas corresponding to the two sub-encoder models respectively include: Wherein, the x m Represents the residual image, where x p This represents the JND point image, where i represents the number of processing steps. This represents the residual image from the previous processing stage that is processed by the multi-level residual compensation network model. The image represents the JND point image processed in the previous step of the multi-level residual compensation network model, where ⊙ represents element-wise multiplication. The step of inputting the JND point image and the residual image into a preset multi-level residual compensation network model and outputting latent representation parameters includes: inputting the residual image into a sub-encoder model for forward transformation processing to obtain a first transformation processing result; performing forward transformation processing on the JND point image and the first transformation processing result through another sub-encoder to obtain a second transformation processing result; and determining the latent representation parameters based on a preset number of processing steps and the second transformation processing result. The quantization and encoding module is used to sequentially quantize and entropy encode the latent representation parameters to obtain the target compressed image corresponding to the original image.

7. An electronic device, characterized in that, The electronic device includes: a memory, a processor, and an image processing program stored in the memory and executable on the processor, wherein the image processing program, when executed by the processor, implements the steps of the image processing method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an image processing program, which, when executed by a processor, implements the steps of the image processing method as described in any one of claims 1 to 5.

Images