A residual mapping method for neural network filters

Abstract

The present invention belongs to the technical field of neural network filters, and is specifically a residual mapping algorithm suitable for neural network filters. The present invention obtains a residual by subtracting the filtering result of the neural network filter from the input of the filter, and adaptively maps the residual to a distortion domain to reduce the size of the reconstruction distortion. The mapping coefficients also need to be encoded and decoded. The present invention reduces distortion and improves the performance of the neural network filter.

Description

Technical Field

[0001] The present invention belongs to the technical field of neural network filters, and in particular relates to a residual mapping method for improving the performance of neural network filters. Background Art

[0002] Neural networks have shown strong performance in the field of image denoising due to their strong nonlinear ability and feature extraction capabilities. Compared with traditional video image compression denoising modules, they have significantly stronger performance. However, neural network filters are often very dependent on training data, and will have more or less overfitting of the training data. How to solve this overfitting phenomenon is a problem that needs to be solved. Effectively solving this overfitting phenomenon can effectively improve the performance of the model.

[0003] The present invention adopts a residual mapping algorithm to map the residual to the distortion domain to achieve a direct change in the model filtering capability, thereby alleviating the occurrence of overfitting. Summary of the invention

[0004] The purpose of the present invention is to propose a residual mapping method suitable for a neural network filter which can alleviate the overfitting phenomenon.

[0005] The residual mapping method for a neural network filter proposed in the present invention adaptively maps the residual to a distortion domain to reduce the size of the reconstruction distortion and improve the performance of the neural network filter. In actual use, it can be used as a supplementary module of the neural network filter and has a certain generalization ability. The specific steps are as follows.

[0006] (1) First, use the neural network filter to filter the reconstructed image to obtain the filtered image; subtract the filtered image from the image before filtering, and define the value obtained as the residual; calculate the difference between the filtered image and the original input image, and define it as the distortion; let R o is distorted, that is, the original pixel Y o The difference between the output X of the deblocking filter and

[0007] R o =Y o -X, (1)

[0008] Similarly, let R S is the residual, that is, the neural network filter output Y S The difference between the output X of the deblocking filter and

[0009] R S =Y S -X, (2).

[0010] (2) Use a function (linear or quadratic) to calculate the residual R S and distortion R o Perform fitting, with the residual as input and the distortion as the target; pre-set function f with λ as parameter λ (·), which can realize R S Therefore, we need to find such a λ that the transformed R S With R o The distance between them is the smallest. For convenience, the distance uses MSE (mean square error); the mean square error is widely used, which can represent the reconstructed PSNR value (peak signal-to-noise ratio) to a considerable extent;

[0011]

[0012] Here, f λ (·) is an example of a linear function (in fact, a complex function, such as a quadratic function, can be used to achieve this), so:

[0013]

[0014] Considering that λ needs to be quantized in practice, we can use the quantized λ to search when executing the search in formula (4), where n is 4;

[0015]

[0016] (3) Use different coefficients in the search to transform the residuals in turn, use the transformed residuals and distortion to calculate the newly generated MSE (mean square error D), and calculate the bit rate R brought by the fitting coefficients. The loss function of the search for the rate-distortion optimization function J is:

[0017] J=D+kR(λ)

[0018] =(λR S -R o ) 2 +kR(λ), (6)

[0019] Among them, the parameter k is the hyperparameter of the rate-distortion optimization function J.

[0020] (4) Among all the candidate functions, save the fitting coefficient with the lowest rate-distortion optimization loss as the final selected coefficient; in fact, because the linear function is selected here, R(λ) is a constant, so when making comparisons, this term can be ignored and only D is considered, that is, the coefficient with the minimum reconstruction loss (λR S -R o ) 2The λ is losslessly entropy coded, and the coding model selects bypass coding. Thus J = D, which simplifies the search process.

[0021] (5) Different from the original method in which the residual before mapping is added to the filter output for reconstruction, here the residual after mapping is added to the filter input for reconstruction, and the fitting coefficients are encoded into the bitstream. Because the fitting coefficients have been encoded into the bitstream, and the same fitting function is preset at the decoding end, λ can be decoded losslessly at the decoding end, thus ensuring the matching of encoding and decoding. Therefore, the reconstructed pixel is It will be used as the input of SAO for subsequent filtering;

[0022] BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Schematic diagram of the method of the present invention.

[0024] Figure 2 Filter position relationship. DETAILED DESCRIPTION

[0025] This section will introduce how to implement the specific residual mapping algorithm. Figure 2 As shown in the figure, CNN-filter (neural network filter) and RM (residual map) should be placed between DB and SAO. The output of DB will not only be connected to CNN-filter, but also be used as input to RM. The output of CNN-filter will also be used as input to RM. The output of RM will eventually be provided to SAO for the next step of filtering.

[0026] (1) For the output of DB, the reconstructed image X is labeled, and the neural network filter and residual mapping algorithm act on the output of DB. The reconstructed image X is filtered using the neural network filter to obtain the filtered image Y S As shown in formula (1), the difference between the image after filtering and the image before filtering is defined as the residual R S As shown in formula (2), the filtered image and the original input image Y are calculated o The difference between o .

[0027] (2) Use a function (linear or quadratic) to fit the residual and distortion, where the residual R S As input, the distortion R o As the goal. There is a requirement for the fitting coefficient, that is, it should be the result of quantization to reduce the consumed bit rate. The quantization rule follows formula (5). In addition to the uniform linear quantization adopted by formula (5), a nonlinear quantization strategy can also be used.

[0028] (3) Use the coefficient λ obtained by fitting to calculate the residual R S Perform the transformation, use the transformed residual and distortion to calculate the newly generated MSE (mean square error), and calculate the rate (bit rate) brought by the fitting coefficient, and use formula (6) to calculate the rate-distortion optimization function J.

[0029] (4) Among all the candidate functions in step (2) and step (3), save the fitting coefficient λ with the lowest rate-distortion optimization loss J as the final selected coefficient.

[0030] (5) Different from the original method, the residual R before mapping S Add to the filter input X as reconstruction, where the mapped residual Added to the filter input X to reconstruct The fitting coefficient λ is subjected to an entropy coding process to complete the coding process.

[0031] (6) At the decoding end, the decoding process needs to decode the output image of DB, that is, the filter input X. Secondly, the residual mapping coefficient λ needs to be decoded. Formula (7) is executed to obtain the mapped residual And add the mapped residual to the filter input X to reconstruct reconstruction It will be sent to the subsequent SAO filter to complete the entire filtering operation.

Claims

1. A residual mapping method suitable for a neural network filter, characterized in that the residual is adaptively mapped to a distortion domain to reduce the size of reconstruction distortion and improve the performance of the neural network filter; the specific steps are as follows: (1) First, use the neural network filter to filter the reconstructed image to obtain the filtered image; subtract the filtered image from the image before filtering, and define the value obtained as the residual; calculate the difference between the filtered image and the original input image, and define it as the distortion; let R o is distorted, that is, the original pixel Y o The difference between the output X of the deblocking filter and R o =Y o -X, (1) Similarly, let R S is the residual, that is, the neural network filter output Y S The difference between the output X of the deblocking filter and R S =Y S -X,(2) (2) Use the function to calculate the residual R S and distortion R o Fitting is performed with the residual as input and the distortion as target; assuming that the function f with λ as parameter λ (·), which can realize R S mapping; we need to find such a λ that the transformed R S With R o The distance between them is the smallest; the distance is measured using the mean square error MSE, which represents the reconstructed PSNR value: Here, for f λ (·) is a linear function, then: Considering that λ needs to be quantized and encoded, when executing the search in formula (4), the quantized λ is used for the search, where n=4: (3) Use different coefficients in the search to transform the residuals in turn, use the transformed residuals and distortion to calculate the newly generated MSE(D), and calculate the bit rate R brought by the fitting coefficients. The loss function of the search for the rate-distortion optimization function J is: J=D+kR(λ) =(λR S -R o ) 2 +kR(λ),(6) in, The parameter k is the hyperparameter of the rate-distortion optimization function J; (4) Among all the candidate functions, the fitting coefficient with the lowest rate-distortion optimization loss is saved as the final selected coefficient; λ (·) is a linear function, so R(λ) is a constant. When making comparisons, we can ignore this term and only consider D, that is, we choose the one with the minimum reconstruction loss (λR S -R o ) 2 The λ is losslessly entropy coded, and the coding model selects bypass coding; thus J = D, which simplifies the search process; (5) The mapped residual is added to the filter input to obtain reconstruction, and the fitting coefficients are encoded into the bitstream. Since the fitting coefficients have been encoded into the bitstream and the same fitting function is preset at the decoding end, λ can be decoded losslessly at the decoding end, thereby ensuring the matching of encoding and decoding. Therefore, the reconstructed pixel is It will be used as the input of SAO for subsequent filtering:

Images