Near-lossless Compressed Image Soft Decoding Method Based on Wide Activation Recurrent Neural Network

Abstract

The present invention belongs to the technical field of digital image processing, and relates to an image soft decoding method, specifically to a near-lossless compressed image soft decoding method based on a wide activation recurrent neural network. The method includes the following steps: S1, acquisition and preprocessing of samples; S2, construction of a wide activation recurrent neural network model; S3, training of the wide activation recurrent neural network model; S4, testing of the wide activation recurrent neural network model. It solves the problems that the quality of the reconstructed compressed image by the existing method is not clear enough and the pixel boundary constraint is not strict enough. The present invention constructs a near-lossless compressed image cyclic reconstruction model based on a wide activation neural network, and redefines the pixel boundary constraint in a reasonable and efficient manner, effectively improving the quality of the reconstructed compressed image, ensuring strict pixel boundary constraints, enhancing the visual sensory experience of people, and can be used to complete the soft decoding of near-lossless compressed images.

Description

Technical Field

[0001] The present invention belongs to the technical field of digital image processing, and relates to an image soft decoding method, in particular to a near-lossless compressed image soft decoding method based on a wide activation recurrent neural network. Background Art

[0002] In the fields of remote sensing images, telemedicine, Internet of Things, etc., image encoders are usually encapsulated in highly integrated chips. Modifying the encoder of the on-chip system to improve image quality is usually a very tricky problem. The advantage of image soft decoding is that it can improve the quality of compressed images without modifying the encoder.

[0003] Soft decoding is actually an ill-posed inverse problem. Previous soft decoding techniques mostly relied on explicit image modeling and optimization. Such as autoregressive models, sparsity-based, random walk graphs, etc. In recent years, with the popularity of convolutional neural networks, researchers have abandoned methods based on hand-designed features and instead designed a data-driven, end-to-end soft decoder with the help of convolutional neural networks. Due to having a larger receptive field and not requiring explicit optimization, soft decoders based on convolutional neural networks are superior to modeling-based soft decoders in terms of image quality improvement and running time. However, most network-based soft decoding techniques only regard the image soft decoding task as a conventional image restoration task. This assumption is not applicable to constrained image restoration tasks such as near-lossless compression.

[0004] Near-lossless compressed image technology is a relatively special image compression technology. In addition to requiring high compression ratios and high image quality, near-lossless compression technology also restricts the peak absolute error of each pixel point in the image. This constraint can preserve tiny local features in the image, so near-lossless compression technology is often used in fields such as remote sensing images and telemedicine that are extremely sensitive to the quality of compressed images.

[0005] Zhou Jiantao et al. published an academic paper titled "L2 Restoration of L ∞ -Decode Images Via Soft-Decision Estimation" in the journal IEEE TRANSACTIONS ON IMAGE PROCESSING, disclosing a near-lossless compressed image soft decoding method based on an autoregressive model. This method improves the quality of low-bitrate image restoration by regularizing quantization errors. However, methods based on autoregressive models usually do not perform as well as convolutional neural networks when dealing with data-driven problems.

[0006] Zhang Xi et al. published an academic paper titled "Near-lossless L" in the journal "Data Compression Conference". ∞ -constrained Image Decompression via Deep Neural Network", which discloses a near-lossless compression soft decoding method based on a generative adversarial network. This method constrains the soft decoding method by adding a loss function with an infinity norm constraint, thereby realizing the recovery of near-lossless compressed images. However, the network used in this method itself has a relatively deep number of layers and a large number of network weight parameters, and can only recover near-lossless compressed images with compression coefficients of 6, 8, and 10. For the soft decoder, it is far from enough to only support the recovery of three compression coefficients. In addition, the loss function with an infinity norm constraint adopted by this method cannot effectively limit the pixel recovery interval of the decoded image.

[0007] In summary, the existing soft decoding methods have problems such as insufficient clarity in reconstructing the quality of compressed images and insufficiently strict pixel boundary constraints. Summary of the Invention

[0008] The purpose of the present invention is to provide a near-lossless compression image soft decoding method based on a wide activation recurrent neural network for the deficiencies in the above-mentioned prior art, so as to solve the problems of insufficient clarity in reconstructing the quality of compressed images and insufficiently strict pixel boundary constraints in the prior art.

[0009] The technical idea of the present invention is: adopting a wide activation convolutional neural network structure, constructing a near-lossless compression soft decoder in a cyclic manner, using a normalized activation function to satisfy the pixel boundary constraints of the recovered image, and completing the recovery of multiple compression coefficients. On the premise of ensuring the improvement of the decoded image quality, by adding a structural infinity norm constraint, the pixel boundary constraints of the near-lossless compression soft decoding method are satisfied, and the decoded image quality is improved.

[0010] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0011] The present application provides a near-lossless compression image soft decoding method based on a wide activation recurrent neural network, and the method includes the following steps: S1, acquisition and preprocessing of samples; S2, construction of a wide activation recurrent neural network model; S3, training of the wide activation recurrent neural network model; S4, testing of the wide activation recurrent neural network model.

[0012] Furthermore, the preprocessing includes performing near-lossless compression on the samples.

[0013] Furthermore, the preprocessing includes dividing the samples into a training sample set and a test sample set at a ratio of 4:1.

[0014] Furthermore, the wide activation recurrent neural network model includes a feature extraction part, a feature inference part, and an image reconstruction part.

[0015] Furthermore, the feature extraction part is an image upscaling network, including an input layer and a convolutional layer.

[0016] Furthermore, the convolutional kernel size of the convolutional layer is 3×3.

[0017] Furthermore, the feature inference part is a wide activation neural network, including an input layer, four convolutional layers, and a residual layer.

[0018] Furthermore, the convolutional kernel size of two of the four convolutional layers is 1×1.

[0019] Furthermore, the activation function of the two convolutional layers is the ReLU activation function.

[0020] Furthermore, the image reconstruction part is an image downscaling network, including an input layer, a convolutional layer, and a residual layer.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] (1) The present invention combines wide activation with a recurrent neural network, greatly reducing the number of network parameters. At the same time, the image reconstruction effect is still better than other similar methods, improving the quality of compressed image reconstruction.

[0023] (2) By introducing a normalized activation function structure, the present invention strictly limits the pixel boundaries of the reconstructed image without the need for additional training functions, effectively improving the quality of compressed image reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a schematic diagram of a near-lossless compressed image soft decoding method based on a wide activation recurrent neural network provided by the present invention;

[0025] Figure 2 is a schematic diagram of the wide activation recurrent neural network constructed in step S2 in a near-lossless compressed image soft decoding method based on a wide activation recurrent neural network provided by the present invention;

[0026] Figure 3 is a comparison chart of the effects of a near-lossless compressed image soft decoding method based on a wide activation recurrent neural network provided by the present invention and other image restoration methods. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] In order to make the implementation process of the present invention clearer, the following will be described in detail with reference to the accompanying drawings.

[0028] The present invention provides a near-lossless compressed image soft decoding method based on a wide activation recurrent neural network, as Figure 1 shown, and the specific steps are as follows:

[0029] S1. Acquisition and preprocessing of samples;

[0030] The image to be restored in the present invention can come from an existing database or be directly captured. Specifically, the images used for training in the present invention come from the existing database DIVK2K, and the images are 900 images of 2K*1K. After training, the present invention can be applied to both database image restoration and direct capture image restoration. The preprocessing includes near-lossless compression, dividing the training sample set and the test sample set, and normalization. M image samples x with a resolution of N×K are taken, and each image sample is compressed using a near-lossless compression method to further obtain M image samples compressed by the near-lossless compression method where M≥900, N≥2000, K≥1000, which is beneficial to constructing a sufficient dataset, accelerating the training convergence of the network, and improving the generalization of the network. Specifically, the number of image samples is not less than 900, which can ensure that the method of the present invention has a sufficient training sample set to learn the mapping relationship between the compressed image samples and the image samples x; the pixels need to be higher than 2000*1000. In addition to ensuring sufficient training, high-resolution image samples contain more detailed information, which helps the method of the present invention learn the restoration of details. More specifically, the near-lossless compression of image samples is a method to achieve compression by reducing the quality of image samples. Therefore, the image samples before compression must ensure high enough resolution and clear enough images to ensure that the neural network can better learn the mapping relationship between the image samples before and after compression, so as to better reconstruct the compressed image. The quantization coefficient is the quantization coefficient for near-lossless compression of the image sample x during near-lossless compression, and the quantization coefficient is an integer. For example, when the quantization coefficient is selected as 2, the near-lossless compression will perform division-by-2 quantization on the prediction error, and the pixel values [0,1,2,3,4] will become [0,0,1,1,2]. Specifically, the larger the quantization coefficient, the more serious the image distortion and the smaller the storage occupied; the smaller the quantization coefficient, the higher the image fidelity, but the less the memory occupied decreases. When the user performs near-lossless compression, it is necessary to comprehensively consider the requirements of memory and fidelity to determine the specific value of the quantization coefficient. Specifically, the near-lossless compression method can refer to the introduction in the academic paper named "Calic-a context based adaptivelossless image codec" published by X. Wu et al. The method of the present invention is applicable to image reconstruction tasks under near-lossless compression with different quantization coefficients, and the applicability of the method of the present invention is relatively strong.

[0031] The image sample x and the corresponding near-lossless compressed image sample are randomly divided into a training sample set and a test sample set at a ratio of 4:1; the training sample set contains 4 / 5 of the image sample x and 4 / 5 of the corresponding near-lossless compressed image sample in it, and the test sample set contains 1 / 5 of the image sample x and 1 / 5 of the corresponding near-lossless compressed image sample in it; since the number of image sample sets constructed by the method of the present invention is large, most of the samples can be used for training and a small part of the samples are left for testing, which not only ensures the sufficiency of the training data but also has a certain amount of data to test the generalization of the method. Normalization is to normalize the image sample to -1 - +1. Specifically, the mean value of the maximum value of the multi-bit image sample is subtracted from the image sample, and then it is divided by the maximum value of the multi-bit image sample. More specifically, in the present invention, the mean value 128 of the 8-bit image maximum value is subtracted from the image sample, and then it is divided by the maximum value 255, and the image normalized to -1 - +1 can be obtained. The mean value 128 and the maximum value 255 are fixed and unchanged, and the processing formula is: Normalizing to -1 - +1 can ensure that the minimum and maximum values of the image are between -1 - +1, so that the error between the reconstructed image obtained by the neural network during training and the image sample x will not be too large, which is convenient for the rapid convergence of training.

[0032] S2. Construction of a wide activation recurrent neural network model;

[0033] S21, as Figure 2 shown, the wide activation recurrent neural network model includes a feature extraction part, a feature inference part, and an image reconstruction part. Among them, the feature extraction part is an image up-dimension network, including an input layer and a convolutional layer, which is used to increase the dimension of the input sample image. In this way, the input image sample can be mapped to another high-dimensional feature space. In addition, higher-dimensional image samples also have more features for subsequent feature extraction; the feature inference part is a wide activation neural network, including an input layer, multiple convolutional layers, and residual layers, which is used to generate the reconstruction image residual; the image reconstruction part is an image down-dimension network, including an input layer, a convolutional layer, and residual layers, which is used to reduce the dimension of the input sample image. In this way, the image dimension can be reduced to be consistent with the output image dimension for subsequent mathematical operations, and at the same time, the input sample image and the reconstruction residual are combined into the finally predicted reconstruction image.

[0034] Specifically, the feature extraction part includes an input layer and a convolutional layer, that is, input layer → first convolutional layer. The feature inference part includes an input layer, four convolutional layers, and a residual layer, which are sequentially set as: input layer → second convolutional layer → third convolutional layer → fourth convolutional layer → fifth convolutional layer → first residual layer. Without changing the size of the image sample, the second convolutional layer further increases the dimension of the image sample. As shown in Table 1, the number of channels of the image sample increases from 64 to 256, so the features that can be extracted by the convolutional operation also increase accordingly. The third convolutional layer reduces the dimension of the image sample. As shown in Table 1, when the number of channels of the image sample is reduced from 256 to 48, that is, the features are screened. In this way, the increase and decrease of the dimension of the image sample by the second convolutional layer and the third convolutional layer respectively enrich and screen the features. The method of the present invention can extract more effective features and improve the quality of the reconstructed image. The fourth convolutional layer (3×1) and the fifth convolutional layer (1×3) are equivalent alternatives to the convolutional kernel (3×3). Their function is to restore the dimension of the image sample to be the same as the input dimension for cyclic inference. The advantage of the equivalent alternative is that although the number of layers increases, the convolutional parameters are reduced from 3×3 to (1×3)×2. The role of the first residual layer is to add the output of the fifth convolutional layer to the input layer, which helps the backpropagation of the training error and facilitates the method to quickly learn. It should be noted that the output of the first residual layer does not directly enter the image reconstruction part, but will return to the input layer of the feature inference part again, and this cycle is performed 12 times. Since it is completely insufficient to use only four convolutional layers for feature inference, but increasing the number of convolutional layers will greatly increase the computational complexity and computational time of the method. To ensure sufficient inference without increasing the method complexity, the method of the present invention uses a cyclic manner for feature inference. The 12 cycles are based on a trade-off between running time and reconstruction effect. If the number of cycles is less than 12, the quality of the image reconstruction will be reduced; if the number of cycles is greater than 12, the reconstruction effect will be improved to a certain extent, but the computational time will also increase accordingly. On the premise of ensuring the quality of the reconstructed image, in order to minimize the running time of the wide activation recurrent neural network model as much as possible, the present invention selects the number of cycles to be 12.

[0035] Specifically, the image reconstruction part includes an input layer, a convolutional layer, and a residual layer, which are arranged in sequence as: input layer → sixth convolutional layer → second residual layer. Similar to the first residual layer in feature inference, the second residual layer adds the output of the sixth convolutional layer to the image to be restored. The role of the sixth convolutional layer is to change the input dimension to be the same as that of the image to be restored. Since the addition operations in the present invention are all pixel-by-pixel additions between images of the same size and the same number of channels, the dimensions of the images must be the same so that the second residual layer can perform the addition operation. Except for the number of output channels of the sixth convolutional layer, the size of the convolutional kernels, the stride, and the number of output channels of other convolutional layers only affect the operation time of the method of the present invention and do not have a great impact on the restoration effect of the image samples. Only the size relationship of the number of output channels needs to be noted, that is, the second convolutional layer should be larger than the first convolutional layer, and the third and fourth convolutional layers should be smaller than the second and fifth convolutional layers to meet the functions of feature enrichment and screening. The output dimension of the sixth convolutional layer needs to be the same as that of the image to be restored, so the number of output channels is fixed at 1.

[0036] Regarding the selection of activation functions for the feature extraction part, the feature inference part, and the image reconstruction part, since the ReLU function is a linear function when the input is positive and outputs zero when the input is negative, and the output of zero means that the gradient corresponding to the input does not need to be calculated, the operation speed is relatively fast. The difference between ELU and ReLU is that when the input is negative, ELU will output a relatively small value, that is, the output is not zero, so that features can also be output when the input is negative, and more features can be extracted. In order to extract more features, the ELU function is selected for the feature extraction part of the present invention. In order to improve the speed of feature inference, the ReLU function is selected for the feature inference part; the image sample is normalized before being input into the wide activation recurrent neural network model. Therefore, before the final output of the wide activation recurrent neural network model, in order to prevent the predicted value range of the wide activation recurrent neural network model from exceeding the range of -1 to +1, the Tanh function is selected for the image reconstruction part, which can force the prediction result to be normalized to -1 to +1 for subsequent inverse transformation operations. Performing one image reconstruction includes performing a feature extraction operation in the feature extraction part once, then entering the feature inference part and looping twelve times. After the inference loop ends, enter the image reconstruction part, and the reconstructed image is output, that is, one image reconstruction is completed. The parameter settings of each convolutional layer in the feature extraction part, the feature inference part, and the image reconstruction part of the present invention are shown in Table 1. The parameter settings in Table 1 are only an example.

[0037] Table 1: Parameter settings of each convolutional layer in the feature extraction part, the feature inference part, and the image reconstruction part.

[0038]

[0039] Such as Figure 2As shown, the quantization coefficient τ is the quantization coefficient for near-lossless compression of the image sample x in step S1. It is 9 integers with a value range of 2 to 10, and needs to be input by the user each time the image sample is restored. For different restorations of image samples, different quantization coefficients τ can be selected to obtain a reconstructed compressed image with higher quality, and the applicability of the method of the present invention is relatively strong. Specifically, each pixel in the image to be restored before inputting into the wide activation recurrent neural network model is divided by the quantization coefficient τ: image / τ; each pixel of the output image sample of the activation function Tanh is multiplied by the quantization coefficient τ: image×τ, which facilitates ensuring the unity of the relational expression during the prediction process, thereby improving the efficiency of reconstructing the image by the method of the present invention.

[0040] S22, define the loss function L of the wide activation recurrent neural network:

[0041]

[0042] where x is the image sample, is the image sample compressed by the near-lossless compression method, F is the wide activation recurrent neural network model, D = H×W×C, H is the height of x and and W is the width of x and and C is the number of channels of x and The present invention uses the two-norm constraint as the loss function and does not need to design a loss function for pixel boundary constraints. The pixel boundary constraints have been realized through the setting of the quantization coefficient τ and the activation function as shown in Figure 2 . Specifically, the image sample x and the compressed image sample satisfy the following relational expression:

[0043]

[0044] where e ∈ [-1, +1]

[0045] where τ is the quantization coefficient and e is the error between the image sample x and the image sample compressed by the near-lossless compression method, that is, it represents the image sample residual. The reconstructed image obtained by the method of the present invention should be consistent with the image sample x, that is, the reconstructed image obtained by the method of the present invention and the image sample to be restored also have the following relational expression:

[0046]

[0047] where e ∈ [-1, +1]

[0048] The method of the present invention uses the Tanh function to represent the error e, so there is the following relational expression:

[0049]

[0050] Among them, is the image reconstruction part and is the input of the Tanh function. Thus, the method of the present invention realizes pixel boundary constraint through function constraint. Compared with the method of designing pixel boundary constraint through loss function, the pixel boundary constraint of the present invention is more strict, can effectively improve the quality of the reconstructed image, ensures strict pixel boundary constraint, and enhances the visual sensory experience of people.

[0051] Table 2: Comparison between the method of the present invention and the L-inftyCNN method.

[0052]

[0053] Specifically, the method of the present invention is compared with the L-inftyCNN method for the results of three quantization coefficients on the LIVE1 dataset. The comparison results are shown in Table 2. Among them, the larger the peak signal-to-noise ratio (PSNR), the smaller the maximum value of the infinity norm boundary represents that the pixel boundary constraint is more strict, and the higher the quality of the reconstructed image. When the quantization coefficients are 6, 8, and 10, the PSNR of the method of the present invention is greater than that of the L-inftyCNN method, and the maximum value of the infinity norm boundary of the method of the present invention is smaller than that of the L-inftyCNN method, that is, the pixel boundary constraint of the method of the present invention is more strict, and the quality of the reconstructed image is higher. To ensure the reliability of the comparison results, the dataset and quantization coefficients used in the method of the present invention are the same as those in this academic paper during the comparison. The specific parameter comparison is shown in Table 2, and the relevant data is from the academic paper titled "Near-lossless `L∞-constrained image decompression via deep neural network" published by X. Zhang et al.

[0054] S3. Training of the wide activation recurrent neural network model;

[0055] Taking the near-lossless compressed image samples in the training sample set as the input of the wide activation recurrent neural network model, and performing E iterations of training on the wide activation recurrent network model to obtain a trained wide activation recurrent network model, where E≥150. The specific training steps are as follows:

[0056] S31. Initializing the parameters of the wide activation neural network, the feature extraction part and the image reconstruction part, setting the current iteration number as T, the maximum iteration number as E, E≥150, and making T = 0.

[0057] S32. Inputting the near-lossless compressed image samples into the feature extraction part to obtain the upsampled image samples The elevation dimension ranges from 1 to 64, mapping the input image samples to another high-dimensional feature space. The high-dimensional sample images have more features for subsequent feature extraction.

[0058] S33. The image samples obtained in step S32 are input into the feature inference part in a cyclic manner for reconstructing image residual prediction. Specifically, the image samples are input into the wide activation neural network to obtain the image sample residuals and the image sample residuals will be input into the feature inference part again to obtain the image sample residuals and are cycled n times in total to obtain the predicted residuals of the reconstructed image samples This can improve the prediction accuracy of the image sample residuals without increasing the method complexity. The number of cycles of this method is 12. Specifically, it corresponds to the number of cycles in step S21. Whenever the feature inference part completes a residual prediction, the image reconstruction part will judge the completion times of the inference part. If it is less than 12 times, the residuals will be sent into the feature inference part again until 12 times are satisfied. Cycling 12 times is a trade-off based on running time and reconstruction effect. If the number of cycles is less than 12, the image reconstruction quality will be reduced; if the number of cycles is greater than 12, the reconstruction effect will be improved to a certain extent, but the operation time will also increase accordingly. On the premise of ensuring the quality of the reconstructed image, in order to minimize the running time of the wide activation recurrent neural network model as much as possible, the present invention selects the number of cycles to be 12.

[0059] S34. The image sample residuals continue to be input into the image reconstruction part to reduce the dimension to 1, that is, to obtain the image sample residuals The image sample residuals after dimensionality reduction after being output by the normalized activation function Tanh, are added to the compressed image samples in step S1 to obtain the final predicted reconstructed image

[0060] S35. The predicted reconstructed image The corresponding image samples x in the training sample set are respectively input into the loss function L, the reconstruction loss after training is calculated, and the gradient descent algorithm is used to update the parameters of the wide activation recurrent neural network through the reconstruction loss. This helps to improve the recovery effect of the method of the present invention, improve the quality of the reconstructed image, and enable the method of the present invention to have better generalization ability, that is, the trained wide activation recurrent neural network model for the feature quantization coefficients can complete the reconstruction of the image samples corresponding to other quantization coefficients, thereby improving the quality of the reconstructed image by the method of the present invention; the gradient descent algorithm helps to generate a stable error gradient, which is convenient for the network to converge, that is, it can quickly find the local optimal solution in the solution space, thereby improving the quality of the reconstructed image by the method of the present invention; when the decrease amplitude of the reconstruction loss is less than 0.0001 or the number of iterations is more than 150 times, stop the training.

[0061] S36, determine whether T is equal to E. If so, obtain the trained wide activation recurrent neural network model; otherwise, let T = T + 1, and restart to execute step S32.

[0062] S4, test of the wide activation recurrent neural network model.

[0063] Use the test sample set obtained in step S1 as the input of the trained wide activation recurrent network model obtained in step S3 to obtain the reconstructed images of the image samples in the test sample set. The PSNR index of the method of the present invention is used as the evaluation score. Generally, the higher the PSNR, the better the recovery effect. During the training process, when the PSNR of the test image no longer increases and this PSNR is higher than that of other similar methods, it is considered that the test is qualified and the trained wide activation recurrent neural network model can be used to recover other images.

[0064] Use the method of the present invention and the image reconstruction methods in related fields to reconstruct the same compressed image data, and the reconstruction results are as Figure 3 shown. Specifically, Figure 3Shows 5 test images from 3 datasets and the corresponding restoration results of different methods. From top to bottom on the left are the 5 test images, and the white boxes are the enlarged detail areas. For the convenience of comparison with existing results, the 5 test images are from common datasets. Among them, the first one from top to bottom on the left is from the Set5 dataset; the second and third ones are from the Kodak PhotoCD dataset; the fourth and fifth ones are from the Manga109 dataset. The smaller images on the right show the detail areas of the restored images of each method.The 10 smaller images corresponding to the test image, from left to right and top to bottom, are the original image (Original), the near-lossless compressed image (NL-CALIC, referring to the academic paper "Calic - a context based adaptive lossless image codec" published by X. Wu et al.), the result of the image restored by the ARCNN method (referring to the academic paper "Compression artifacts reduction by a deep convolutional network" published by C. Dong et al.), the result of the image restored by the SRCNN method (referring to the academic paper "Image super-resolution using deep convolutional networks" published by C. Dong et al.), the result of the image restored by the FSRCNN method (referring to the academic paper "Accelerating the super-resolution convolutional neural network" published by C. Dong et al.), the result of the image restored by the ShuffleNet method (referring to the academic paper "Shufflenet: An extremely efficient convolutional neural network for mobile devices" published by X. Zhang et al.), the result of the image restored by the GhostNet method (referring to the academic paper "Ghostnet: More features from cheap operations" published by K. Han et al.), the result of the image restored by the PAN method (referring to the academic paper "Efficient image super-resolution using pixel attention" published by H. Zhao et al.), the result of the image restored by the MIRNet method (referring to the academic paper "Learning enriched features for real image restoration and enhancement" published by S. W. Zamir et al.), and the result of the image restored by the method of the present invention (WRSD); among them, the ARCNN, SRCNN, FSRCNN, ShuffleNet, GhostNet, and PAN methods all belong to image restoration methods, and the MIRNet method belongs to the method of restoring images by a reference model.Since the number of parameters of MIRNet is much larger than that of this method (600 times), and the quality of the restored images is better than that of other comparison methods, the results of the MIRNet method are mainly used as a reference. Figure 3 From the comparison of the 5 sets of smaller images in Figure 3 , it can be seen that the images restored by the method of the present invention are smoother and more natural than the restored results obtained by other comparison methods. Therefore, the compressed images reconstructed according to the method of the present invention are clearer and the quality of the reconstructed images is higher.

[0065] The hardware test platform adopted in this application is: an Intel Core i7 CPU with a main frequency of 3.60 GHz and 8 GB of memory; the software simulation platform is: Ubuntu 16.04 64-bit operating system, Pycharm development platform; the software simulation language: Python; the deep learning framework used: Pytorch.

[0066] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A near-lossless compressed image soft decoding method based on a wide activation recurrent neural network, characterized in that The method includes the following steps: S1, acquisition and preprocessing of samples; S2, construction of a wide activation recurrent neural network model; S3, training of the wide activation recurrent neural network model; S4, testing of the wide activation recurrent neural network model; The wide activation recurrent neural network model includes a feature extraction part, a feature inference part, and an image reconstruction part; the feature extraction part is an image upsampling network, including an input layer and a first convolutional layer, and the convolutional kernel size of the first convolutional layer is 3×3; The feature inference part is a wide activation neural network, including an input layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, and a first residual layer; the convolutional kernel sizes of the second convolutional layer and the third convolutional layer are 1×1; the activation functions of the second convolutional layer and the third convolutional layer are both ReLU activation functions, and the image reconstruction part is an image downsampling network, including an input layer, a sixth convolutional layer, and a second residual layer; The loss function L of the wide activation recurrent neural network is: where x is an image sample, is the image sample compressed using a near-lossless compression method, F is a wide activation recurrent neural network model, D = H × W × C, H is the height of x and , W is the width of x and , C is the number of channels of x and ; the following relationship holds between x and : where e ∈ [-1, +1], τ is the quantization coefficient, and e is the image sample residual The step S3 includes the following steps: S31, initialize the parameters of the wide activation neural network, the feature extraction part, and the image reconstruction part, set the current iteration number as T, the maximum iteration number as E, E≥150, and let T = 0; S32, near-lossless compression of the image sample Input the image sample into the feature extraction part to obtain the image sample after dimensionality increase The dimensionality is increased from 1 to 64, and the input image sample is mapped to a high-dimensional feature space S33. Input the image sample obtained in the step S32 into the feature inference part in a cyclic manner to predict the reconstructed image residual; the image sample After being input into the wide activation neural network, an image sample residual is obtained The image sample residual Is input into the feature inference part again to obtain the image sample residual A total of n cycles are performed to obtain the predicted residual of the reconstructed image sample S34, Image sample residual Input into the image reconstruction part to reduce the dimension to 1, i.e., obtain the image sample residual The image sample residual after dimensionality reduction After being output by the normalized activation function Tanh, it is added to the compressed image sample in step S1 to obtain the final predicted reconstructed image S35, bring the predicted reconstructed image and the corresponding image sample in the training sample set into the loss function L respectively, calculate the reconstruction loss after training, and use the gradient descent algorithm to update the parameters of the wide activation recurrent neural network through the reconstruction loss; stop training when the decrease amplitude of the reconstruction loss is less than 0.0001 or the number of iterations is more than 150 times; S36, determine whether T is equal to E, if so, obtain the trained wide activation recurrent neural network model; otherwise, let T = T + 1, and restart the execution of the step S32.

2. The near-lossless compressed image soft decoding method based on a wide activation recurrent neural network according to claim 1, wherein The preprocessing includes performing near-lossless compression on the samples.

3. The near-lossless compressed image soft decoding method based on a wide activation recurrent neural network according to claim 2, wherein The preprocessing includes dividing the samples into a training sample set and a test sample set at a ratio of 4:1.

Images