Pig image BDE reconstruction system and method based on differential image rate filtering

Abstract

The invention relates to a pig image BDE reconstruction system and method based on differentiatable image rate filtering, and belongs to the field of image processing, and the system comprises an image preprocessing module which is configured to process an original low-bit image; the neural network module acts on the processing and reconstruction tasks of low-bit pig images, the characteristics of each image are collaboratively optimized in a task dividing and conquering mode, and finally a high-quality image is obtained; the loss function module is configured to optimize the network model performance by minimizing the difference between the network output image and the real image; the training module is configured to optimize network model parameters and a learning process by adopting a staged training strategy; and the test and evaluation module is configured to evaluate the image quality through the peak signal-to-noise ratio and the structural similarity index and measure the reconstruction effect of the model. According to the method, a two-stage training strategy is adopted, an image reconstruction task is decomposed into two sub-tasks which cooperate with each other, and spatial resolution enhancement and bit resolution improvement are achieved.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a pig image BDE reconstruction system and method based on differentiable image rate filtering. Background Art

[0002] Differentiable image rate filtering technology is a new image processing method that combines traditional frequency domain processing with deep learning frameworks in recent years. Its core lies in constructing the frequency domain filtering process as a differentiable operator, so as to be compatible with gradient-based end-to-end optimization. Unlike traditional fixed-parameter frequency domain filters (such as ideal low-pass filters), this method dynamically generates frequency response functions through neural networks, and performs derivable amplitude modulation or phase adjustment on the image spectrum in the Fourier transform domain. It can not only maintain the advantages of traditional frequency domain processing in noise separation and feature extraction, but also automatically optimize the filtering characteristics through back propagation. This technology breaks through the limitations of manually designed filter parameters in traditional methods, and shows stronger adaptability and detail preservation capabilities in tasks such as image deblurring and high dynamic range imaging. Especially when dealing with non-stationary noise and complex degradation scenes, the optimal frequency domain feature mapping relationship can be learned in a data-driven manner.

[0003] With the rapid development of the breeding industry, the demand for intelligent breeding management systems is increasing. Among them, pig vital signs monitoring, as an important part of breeding management, relies on high-precision image processing technology. However, in actual breeding environments, due to factors such as lighting changes, haze interference, and motion blur, pig images often have quality degradation problems, making it difficult to clearly identify key features (such as ear tags, hair textures, contours, etc.). Traditional image processing methods perform poorly in dealing with these complex degradation scenarios and are difficult to meet the high requirements for image quality and processing efficiency in breeding management. Traditional methods usually use fixed parameter filters (such as mean filtering, Gaussian filtering) or rule-based image enhancement techniques, which have obvious limitations when processing pig images.

[0004] The defects of existing traditional methods are as follows: 1. The limitations of fixed parameter filters are: some artifact removal algorithms (such as CRR, CA, FACE-BDE, BEF-BDE) rely on manually designed filter parameters, lack flexibility and adaptability, and cannot be dynamically adjusted according to the local characteristics of the image (such as brightness distribution, color gamut offset, degradation type), resulting in poor results when processing complex degradation of pig images. For example, fixed parameter filters are prone to over-smoothing when processing low-light or haze images, resulting in the loss of high-frequency details (such as ear tags, hair texture); and when processing motion blurred images, they cannot effectively suppress low-frequency artifacts (such as abdominal color blocks). 2. The limitations of fixed parameter filters are: some artifact removal algorithms (such as CRR, CA, FACE-BDE, BEF-BDE) rely on manually designed filter parameters, lack flexibility and adaptability, and cannot be dynamically adjusted according to the local characteristics of the image (such as brightness distribution, color gamut offset, degradation type), resulting in poor results when processing complex degradation of pig images. For example, fixed parameter filters are prone to over-smoothing when processing low-light or haze images, resulting in the loss of high-frequency details (such as ear tags and hair textures); and when processing motion-blurred images, they cannot effectively suppress low-frequency artifacts (such as abdominal color blocks). 3. The limitations of the gradient conduction path are: the existing methods lack a dynamic trade-off between low-frequency suppression and high-frequency enhancement during the gradient optimization process, which makes it difficult for the model to find the optimal balance point during the training process. For example, when denoising, conventional methods may smooth the entire image, but at the same time weaken the edge details, resulting in blurred target features; while when enhancing the edges, artifacts may be amplified, leaving color blocks. Therefore, the traditional gradient optimization strategy is not smart enough in adjusting the weights between low-frequency and high-frequency information, resulting in unstable final restoration effects. 4. The limitations of the lack of a joint optimization mechanism are: some methods attempt to divide artifact removal and detail restoration into different stages, but they are usually optimized independently, lack information interaction across stages, and cannot achieve a dynamic balance between low-frequency suppression and high-frequency enhancement. 5. The shortcomings of multi-scale feature fusion: existing methods mostly use single-scale processing, which makes it difficult to simultaneously restore the global structure and local details of the image. For example, when processing high-resolution pig images, traditional methods cannot effectively utilize multi-scale information, resulting in loss of image details or distortion of global structures. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a pig image BDE reconstruction system and method based on differentiable image rate filtering to solve the shortcomings of the prior art.

[0006] The object of the present invention is achieved by the following technical solutions: a pig image BDE reconstruction system based on differentiable image rate filtering, which includes an image preprocessing module, a neural network module, a loss function module, a training module and a testing and evaluation module; The image preprocessing module is configured to process the original low-bit image; The neural network module includes a spatial resolution enhancement submodule, a bit resolution enhancement submodule and an adaptive submodule. The three submodules work together to process and reconstruct low-bit pig images, and coordinately optimize each image feature by task division, ultimately obtaining a high-quality image. The loss function module is configured to optimize the network model performance by minimizing the difference between the network output image and the real image; The training module is configured to optimize the network model parameters and learning process by adopting a phased training strategy to reduce interference between tasks, enhance the synergy between modules, and improve the scalability and adaptability of the fusion network model; The testing and evaluation module is configured to evaluate image quality through peak signal-to-noise ratio and structural similarity index, and measure the reconstruction effect of the model.

[0007] The spatial resolution enhancement submodule is used to improve the spatial resolution of the input image, restore the details of the image and reduce the blur and detail loss caused by low resolution; The spatial resolution enhancement submodule includes three parts: convolution layer, residual block and global residual learning. The input low spatial resolution and low bit pig image is used to extract preliminary features through the convolution layer, and the negative semi-axis slope is automatically adjusted through training to improve the nonlinear expression ability. Then the residual block maintains the pig image features and avoids gradient disappearance through jump connection. Finally, global residual learning is performed and finally added to the input low resolution and low bit pig image to restore image details.

[0008] The bit resolution enhancement submodule is used to restore details in low-bit images and eliminate pseudo contour problems caused by low-bit quantization; The network architecture of the bit resolution enhancement submodule is the same as that of the spatial resolution enhancement submodule, both of which include three parts: convolutional layer, residual block and global residual learning. The bit resolution enhancement submodule focuses on improving the bit resolution of low spatial resolution and low-bit pig images, and uses the perceptual loss function to repair the artifacts and color distortion caused by low-bit quantization.

[0009] The adaptive submodule performs weighted fusion on the images output from the spatial resolution enhancement submodule and the bit resolution enhancement submodule by learning an adaptive fusion mask to obtain a final high-quality image; The adaptive submodule includes three parts: an input layer, a residual block and an output layer. The input layer receives low spatial resolution and low bit pig images and passes them to the subsequent convolutional layer for feature extraction. The residual block is then used to enhance the expression ability of image features, and information loss is avoided through jump connections. Finally, the output layer generates a fusion mask to control the fusion ratio of the output images of the spatial resolution enhancement submodule and the bit resolution enhancement submodule to obtain the final high-quality pig image.

[0010] A pig image BDE reconstruction method based on differentiable image rate filtering, the method comprising: Image preprocessing step: Process the original low-bit image to provide standardized input data for subsequent training; Training steps: Use the spatial resolution enhancement submodule to perform spatial resolution enhancement training, use the bit resolution enhancement submodule to perform bit resolution enhancement training, update the network model parameters and optimize the network model performance through the loss function, and use the adaptive submodule to adaptively combine the outputs of the spatial resolution enhancement submodule and the bit resolution enhancement submodule by learning the fusion mask to obtain the final image, and freeze all network model parameters to obtain a trained network model; Reasoning steps: input low spatial resolution and low bit pig images to the spatial resolution enhancement submodule and the bit resolution enhancement submodule, output high spatial resolution and low bit pig images and low spatial resolution and high bit resolution pig images, the adaptive submodule outputs a fusion mask according to the input low spatial resolution and low bit pig images to control the fusion ratio of the spatial resolution enhancement pig images and the bit resolution enhancement pig images to obtain high spatial resolution and high bit pig images.

[0011] The training step specifically includes a spatial resolution enhancement submodule training step, a bit resolution enhancement submodule training step and an adaptive submodule training step; The spatial resolution enhancement submodule training includes: inputting a low spatial resolution and low bit pig image, restoring the spatial resolution of the image through the spatial resolution enhancement submodule, using the mean square error to optimize the detail recovery of the image to ensure that the enhanced image has high spatial resolution and rich details, and using the Adam optimizer combined with the learning rate adjustment strategy to improve the convergence speed and avoid gradient disappearance, and finally generating a high spatial resolution and low bit pig image.

[0012] The bit resolution enhancement submodule training includes: inputting a low spatial resolution and low bit pig image, eliminating pseudo contours and quantization noise in the low spatial resolution and low bit pig image through the bit resolution enhancement submodule, using a perceptual loss function to ensure that the bit depth of the output image is close to the real image and reduce the pseudo contour phenomenon, and using an Adam optimizer to ensure that network parameters are updated during the training process to improve the bit depth quality, and finally generate a low resolution and high bit pig image.

[0013] The adaptive submodule training includes: inputting a low spatial resolution and low bit pig image, adaptively combining the outputs of the spatial resolution enhancement submodule and the bit resolution enhancement submodule by learning a fusion mask to obtain a final fused image, using a mean absolute error or a perceptual loss function to calculate the difference between the fused image and a true high bit and high spatial resolution image, and using an Adam optimizer combined with a cosine annealing schedule to dynamically adjust the learning rate to improve the stability of the training, and finally obtaining a high spatial resolution and high bit pig image.

[0014] The method also includes a testing and evaluation step: evaluating the image quality through peak signal-to-noise ratio and structural similarity index to measure the reconstruction effect of the model.

[0015] The peak signal-to-noise ratio is used to reflect the error level of image restoration, and the higher the value, the better the image quality; the structural similarity index is used to evaluate the consistency of image structural information to be closer to the quality perceived by the human eye.

[0016] The present invention has the following advantages: 1. A two-stage training strategy is adopted to decompose the image reconstruction task into two mutually coordinated subtasks to achieve spatial resolution enhancement and bit resolution improvement. This non-independent task decomposition method enables each sub-network to give full play to its own advantages and avoid the problem of mutual interference between low-frequency and high-frequency processing in traditional methods through cross-module information interaction, thereby significantly improving the overall reconstruction effect and system stability.

[0017] 2. By adopting the differentiated design of SRNet and BDENet, and using the adaptively learned fusion mask in the fusion stage, the entire network structure can not only flexibly cope with different types of degradation (such as low light, motion blur, haze interference, etc.), but also realize joint optimization of each module in the back propagation process. In this way, not only can the pseudo contours caused by low-bit quantization and the detail loss caused by low resolution be dynamically adjusted, but also the restoration of global structure and local details can be taken into account, ultimately achieving a better image reconstruction effect.

[0018] 3. After training SRNet and BDENet independently in stages, FusionNet is used to fuse the two results, and dynamic interaction of cross-stage information is achieved through gradient back propagation. At the same time, progressive optimization strategies (such as cosine annealing scheduling, Adam optimizer, etc.) are used to ensure stable convergence during training, and gradually improve the network's adaptability to complex degradation scenarios at different training stages. This dynamic joint training strategy not only improves the model's ability to restore details and colors in low-bit image reconstruction, but also greatly improves the real-time and robustness of the system in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the process of the pig image BDE reconstruction method of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application provided below in conjunction with the drawings is not intended to limit the scope of protection of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application. The present invention is further described below in conjunction with the drawings.

[0021] The present invention aims to solve the problems of insufficient collaborative optimization, low efficiency and unsatisfactory reconstruction quality in the existing pig image reconstruction technology. One implementation method proposes a pig image reconstruction system based on differentiable image rate filtering and bilateral diffusion equation (BDE), which includes the following contents.

[0022] 1. Neural network module: It includes three core modules: spatial resolution enhancement submodule (SRNet), bit resolution enhancement submodule (BDENet) and adaptive fusion submodule (FusionNet). These submodules work together to process and reconstruct low spatial resolution and low bit pig images, and optimize each image feature through task division and finally obtain high-quality images.

[0023] Among them, the spatial resolution enhancement submodule is used to improve the spatial resolution of the input low spatial resolution and low bit pig image, restore the details of the image and reduce the blur and detail loss caused by low resolution, and output a high spatial resolution and low bit pig image. It includes three parts: convolution layer, residual block and global residual learning.

[0024] Furthermore, the convolution layer: The first convolution layer uses a 3x3 convolution kernel, a step size of 1, and an output channel of 64 to extract preliminary features. The activation function uses PReLU (Parametric ReLU), and the negative semi-axis slope is automatically adjusted through training to improve the nonlinear expression ability.

[0025] Residual block: SRNet contains 16 residual blocks, each of which consists of two 3x3 convolutional layers and a PReLU activation function. Each residual block uses a skip connection to preserve the original image features and avoid the gradient vanishing problem.

[0026] Global residual learning: The output after 16 residual blocks is passed through a 3x3 convolutional layer for global residual learning, and finally added to the input pig image to restore image details. The activation function is Tanh, which normalizes the pixel value to the range of [-1, 1] to adapt to image restoration.

[0027] Among them, the bit resolution enhancement submodule is used to restore the details in the low-spatial-resolution and low-bit pig images and eliminate the pseudo contour problem caused by low-bit quantization. BDENet has the same network architecture as SRNet, but its processing method focuses on improving the bit resolution. Therefore, during the training process, the network parameters are updated and optimized in a different way from SRNet. The perceptual loss function is used to repair the artifacts and color distortion caused by low-bit quantization, and finally outputs low-spatial-resolution and high-bit pig images.

[0028] Among them, the adaptive fusion submodule learns the adaptive fusion mask , weighted fusion of images output from the spatial resolution enhancement submodule (SRNet) and the bit resolution enhancement submodule (BDENet) is performed to obtain the final high-quality image. This module only performs image fusion operations and does not involve complex mappings such as blurring or pseudo contour suppression. The fusion coefficients are adaptively learned through the network and are guaranteed to be non-negative and sum to 1. It consists of three parts: input layer, residual block, and output layer.

[0029] Furthermore, the input layer receives low-bit and low-spatial-resolution pig images and passes them to the subsequent convolutional layers for feature extraction.

[0030] Residual block: Contains 16 residual blocks, each of which consists of two 3x3 convolutional layers and a PReLU activation function. Residual learning is used to enhance the expressiveness of image features. Skip connections help avoid information loss.

[0031] Output layer: Generate fusion mask , control high spatial resolution low bit pig images and low spatial resolution high bit pig images The fusion ratio of the mask value is between [0, 1], and the Sigmoid activation function is used to ensure that the fusion coefficient is non-negative and the sum is 1. The fusion formula is .

[0032] 2. Loss function module: Its function is to optimize the performance of the model by minimizing the difference between the network output image and the real image.

[0033] Among them, the spatial resolution enhancement loss function ( )for: , by comparing the output high spatial resolution low bit pig image High spatial resolution and real high bit pig image The difference is calculated using the mean square error (MSE) to optimize the network.

[0034] Bit resolution enhancement loss function ( )for: , the loss function is calculated by calculating the low spatial resolution high bit pig image High spatial resolution and real high bit pig image The difference between the two is analyzed and BDENet is optimized to eliminate the pseudo contour problem.

[0035] The fusion loss function is: When training FusionNet, the fusion loss is minimized and the fusion mask is optimized to make the final image closer to the real image. The loss function is the mean square error (MSE).

[0036] 3. Training module: responsible for the parameter optimization and learning process of the neural network of the present invention. Its core goal is to enable the network to output high-quality, high-resolution images when inputting low-bit pig images. The entire training process adopts a phased training strategy to reduce interference between tasks, enhance the synergy between modules, and improve the scalability and adaptability of the fusion network.

[0037] The training process consists of the following two stages: Stage 1: Independent training of SRNet and BDENet: Input low spatial resolution and low bit pig images , output high spatial resolution low bit pig image and low spatial resolution high bit pig images .

[0038] Stage 2: Input low spatial resolution low bit pig image , output the fusion mask generated by FusionNet The final fused image .

[0039] 4. Test evaluation module: Evaluate image quality through PSNR (peak signal-to-noise ratio) and SSIM (structural similarity index) to measure the reconstruction effect of the network model.

[0040] Among them, the PSNR calculation formula is: ; in, Represents the square of the maximum possible value of the image pixel. For common 8-bit images, the pixel value range is 0 to 255. = 255, so the numerator is 255 2 , for images with different bit depths (such as 16 bits), the maximum value will be adjusted accordingly (such as 65535), and the numerator represents the maximum theoretical intensity of the signal. MSE stands for mean square error, which is the square average of the difference between each pixel of the original image and the processed image.

[0041] SSIM calculation formula: ; Among them, x and y are two images to be compared, and represents the average value of image x and y, and represents the variance of images x and y, represents the covariance of images x and y, and Represents a small constant introduced to avoid the denominator being zero.

[0042] like Figure 1 As shown, another embodiment of the present invention relates to a pig image reconstruction method based on differentiable image rate filtering and bilateral diffusion equation (BDE), which specifically includes the following contents.

[0043] S1. Image preprocessing step: First, the original low-bit image is processed to provide standardized input for subsequent training, including scale normalization and pixel value normalization of low spatial resolution and low-bit images.

[0044] Among them, scale normalization includes: using bilinear interpolation to adjust the image size to the target size (such as 256×256) to ensure that the image size is consistent for subsequent network processing.

[0045] Pixel value normalization includes: normalizing the image pixel values ​​to the range of [0,1] or [-1,1] to adapt to the input requirements of the deep learning model and ensure the training stability of the network.

[0046] S2, training steps; Phase 1: Independent training of SRNet and BDENet: In this phase, the task is decomposed into two subtasks, using SRNet and BDENet for training spatial resolution enhancement and bit resolution enhancement respectively.

[0047] SRNet training includes: Input low spatial resolution low bit pig image , restore the spatial resolution of the image through SRNet, and generate an enhanced high spatial resolution low bit pig image The loss function uses the mean square error (MSE) to optimize the detail recovery of the image, ensuring that the enhanced image has high spatial resolution and rich details. The optimizer uses the Adam optimizer, combined with the learning rate adjustment strategy, to improve the convergence speed and avoid gradient disappearance.

[0048] BDENet training involves: Input low spatial resolution low bit pig image , BDENet is used to eliminate pseudo contours and quantization noise in low spatial resolution and low bit pig images, and generate low spatial resolution and high bit images The loss function uses the perceptual loss function to ensure that the bit depth of the output image is close to the real image and reduce the pseudo contour phenomenon. The optimizer uses the Adam optimizer to ensure that the network parameters are updated during the training process to improve the bit depth quality.

[0049] Phase 2: After SRNet and BDENet are trained independently, FusionNet begins training. The main goal of this phase is to adaptively combine the outputs of SRNet and BDENet by learning the fusion mask to obtain the final fused image.

[0050] Input low spatial resolution low bit pig image , the image is used as the input of SRNet and BDENet at the same time, by learning a fusion mask , controls the weighted ratio of SRNet and BDENet outputs, and the fusion formula is: FusionNet adaptively adjusts the fusion ratio according to the image features and outputs the final image ; The loss function uses mean square error (MSE) or perceptual loss to calculate the difference between the fused image and the real high-bit high spatial resolution image The training focus at this stage is to learn the optimal fusion mask The optimization strategy uses the Adam optimizer combined with cosine annealing to dynamically adjust the learning rate to improve the stability of training. It only updates the weight parameters of FusionNet to ensure stability during training.

[0051] After the training step is completed, all network parameters are frozen and no updates are performed.

[0052] S3, inference step: is the process used for practical application after the network model training is completed. The main difference is that no weight update is performed and all network parameters are fixed. The goal of the inference stage is to generate high-quality images in real time.

[0053] Input image: Input low-bit and low-spatial-resolution pig image , the image is processed by the trained network model to obtain the final reconstructed image.

[0054] Network model reasoning: SRNet receives input low spatial resolution low bit pig image , output spatial resolution enhanced pig image ; BDENet receives the same input low spatial resolution low bit pig image , output bit resolution enhanced pig image FusionNet is based on the input low spatial resolution and low bit pig image , output fusion mask , used to control the fusion ratio of SRNet and BDENet outputs.

[0055] Image fusion: Fusion mask generated using FusionNet , the outputs of SRNet and BDENet are weighted and fused according to the following formula: , This process ensures the effective combination of spatial resolution enhancement and bit resolution enhancement tasks, and the final output image has higher spatial resolution, bit depth and image details.

[0056] The process of the inference step is end-to-end, and all processing steps can be completed with only one forward propagation. There is no need to retrain the model when it is deployed, and the input image can quickly generate the output image.

[0057] S4. Testing and evaluation steps: The image quality is evaluated by PSNR (peak signal-to-noise ratio) and SSIM (structural similarity index) to measure the reconstruction effect of the network model.

[0058] PSNR (Peak Signal-to-Noise Ratio): reflects the error level of image restoration. The higher the value, the better the image quality. SSIM (Structural Similarity Index): Evaluates the consistency of image structural information, which is closer to the quality perceived by the human eye.

[0059] Through comprehensive analysis of these two indicators, the image restoration quality and model performance are quantitatively evaluated.

[0060] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used for various other combinations, modifications and improvements, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not deviate from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A pig image BDE reconstruction system based on differentiable image rate filtering, characterized by: It includes image preprocessing module, neural network module, loss function module, training module and testing and evaluation module; The image preprocessing module is configured to process the original low-bit image; The neural network module includes a spatial resolution enhancement submodule, a bit resolution enhancement submodule and an adaptive submodule. The three submodules work together to process and reconstruct low-bit pig images, and coordinately optimize each image feature by task division, ultimately obtaining a high-quality image. The loss function module is configured to optimize the network model performance by minimizing the difference between the network output image and the real image; The training module is configured to optimize the network model parameters and learning process by adopting a phased training strategy to reduce interference between tasks, enhance the synergy between modules, and improve the scalability and adaptability of the fusion network model; The testing and evaluation module is configured to evaluate image quality through peak signal-to-noise ratio and structural similarity index, and measure the reconstruction effect of the model.

2. The pig image BDE reconstruction system based on differentiable image rate filtering according to claim 1, characterized in that: The spatial resolution enhancement submodule is used to improve the spatial resolution of the input image, restore the details of the image and reduce the blur and detail loss caused by low resolution; The spatial resolution enhancement submodule includes three parts: convolution layer, residual block and global residual learning. The input low spatial resolution and low bit pig image is used to extract preliminary features through the convolution layer, and the negative semi-axis slope is automatically adjusted through training to improve the nonlinear expression ability. Then the residual block maintains the pig image features and avoids gradient disappearance through jump connection. Finally, global residual learning is performed and finally added to the input low resolution and low bit pig image to restore image details.

3. The pig image BDE reconstruction system based on differentiable image rate filtering according to claim 1, characterized in that: The bit resolution enhancement submodule is used to restore details in low-bit images and eliminate pseudo contour problems caused by low-bit quantization; The network architecture of the bit resolution enhancement submodule is the same as that of the spatial resolution enhancement submodule, both of which include three parts: convolutional layer, residual block and global residual learning. The bit resolution enhancement submodule focuses on improving the bit resolution of low spatial resolution and low-bit pig images, and uses the perceptual loss function to repair the artifacts and color distortion caused by low-bit quantization.

4. The pig image BDE reconstruction system based on differentiable image rate filtering according to claim 1, characterized in that: The adaptive submodule performs weighted fusion on the images output from the spatial resolution enhancement submodule and the bit resolution enhancement submodule by learning an adaptive fusion mask to obtain a final high-quality image; The adaptive submodule includes three parts: an input layer, a residual block and an output layer. The input layer receives low spatial resolution and low bit pig images and passes them to the subsequent convolutional layer for feature extraction. The residual block is then used to enhance the expression ability of image features, and information loss is avoided through jump connections. Finally, the output layer generates a fusion mask to control the fusion ratio of the output images of the spatial resolution enhancement submodule and the bit resolution enhancement submodule to obtain the final high-quality pig image.

5. A pig image BDE reconstruction method based on differentiable image rate filtering, characterized by: The pig image BDE reconstruction method comprises: Image preprocessing step: Process the original low-bit image to provide standardized input data for subsequent training; Training steps: Use the spatial resolution enhancement submodule to perform spatial resolution enhancement training, use the bit resolution enhancement submodule to perform bit resolution enhancement training, update the network model parameters and optimize the network model performance through the loss function, and use the adaptive submodule to adaptively combine the outputs of the spatial resolution enhancement submodule and the bit resolution enhancement submodule by learning the fusion mask to obtain the final image, and freeze all network model parameters to obtain a trained network model; Reasoning steps: input low spatial resolution and low bit pig images to the spatial resolution enhancement submodule and the bit resolution enhancement submodule, output high spatial resolution and low bit pig images and low spatial resolution and high bit resolution pig images, the adaptive submodule outputs a fusion mask according to the input low spatial resolution and low bit pig images to control the fusion ratio of the spatial resolution enhancement pig images and the bit resolution enhancement pig images to obtain high spatial resolution and high bit pig images.

6. The pig image BDE reconstruction method based on differentiable image rate filtering according to claim 5, characterized in that: The training step specifically includes a spatial resolution enhancement submodule training step, a bit resolution enhancement submodule training step and an adaptive submodule training step; The spatial resolution enhancement submodule training includes: inputting a low spatial resolution and low bit pig image, restoring the spatial resolution of the image through the spatial resolution enhancement submodule, using the mean square error to optimize the detail recovery of the image to ensure that the enhanced image has high spatial resolution and rich details, and using the Adam optimizer combined with the learning rate adjustment strategy to improve the convergence speed and avoid gradient disappearance, and finally generating a high spatial resolution and low bit pig image.

7. The pig image BDE reconstruction method based on differentiable image rate filtering according to claim 6, characterized in that: The bit resolution enhancement submodule training includes: inputting a low spatial resolution and low bit pig image, eliminating pseudo contours and quantization noise in the low spatial resolution and low bit pig image through the bit resolution enhancement submodule, using a perceptual loss function to ensure that the bit depth of the output image is close to the real image and reduce the pseudo contour phenomenon, and using an Adam optimizer to ensure that network parameters are updated during the training process to improve the bit depth quality, and finally generate a low resolution and high bit pig image.

8. The pig image BDE reconstruction method based on differentiable image rate filtering according to claim 7, characterized in that: The adaptive submodule training includes: inputting a low spatial resolution and low bit pig image, adaptively combining the outputs of the spatial resolution enhancement submodule and the bit resolution enhancement submodule by learning a fusion mask to obtain a final fused image, using a mean absolute error or a perceptual loss function to calculate the difference between the fused image and a true high bit and high spatial resolution image, and using an Adam optimizer combined with a cosine annealing schedule to dynamically adjust the learning rate to improve the stability of the training, and finally obtaining a high spatial resolution and high bit pig image.

9. The pig image BDE reconstruction method based on differentiable image rate filtering according to claim 5, characterized in that: The method also includes a testing and evaluation step: evaluating the image quality through peak signal-to-noise ratio and structural similarity index to measure the reconstruction effect of the model.

10. The pig image BDE reconstruction method based on differentiable image rate filtering according to claim 9, characterized in that: The peak signal-to-noise ratio is used to reflect the error level of image restoration, and the higher the value, the better the image quality; the structural similarity index is used to evaluate the consistency of image structural information to be closer to the quality perceived by the human eye.

Images