A method for training a noise reduction neural network and a noise reduction method

Abstract

The application provides a denoising neural network training method and a denoising method. The denoising neural network comprises n denoising units, and each denoising unit is trained separately. The training of the first denoising unit comprises the following steps: taking an image without noise as a base image; adding first noise to the base image to obtain a first noise image; removing part of the first noise to obtain second noise, and adding the second noise to the base image to obtain a second noise image; taking the first noise image as a training input image and taking the second noise image as a training output image; training the first denoising unit to obtain a trained first denoising unit. The second noise image still has part of the noise. The training method can control the denoising strength of the denoising unit. After that, the network with different denoising strengths can be provided by combining multiple denoising units, and it is not necessary to retrain different networks to change the denoising strength.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a denoising neural network training method and a denoising method. BACKGROUND

[0002] Image denoising refers to removing noise in an image by using a specific processing method to estimate and restore the true signal value of the image. In recent years, the use of neural network training for image denoising has become more and more common. The advantage of this technology is that it overcomes some shortcomings of traditional artificially designed denoising algorithms, such as image blurring after denoising and insufficient denoising effect. The training of a neural network relies on a large amount of training data sets and a complex model structure. Such complexity usually makes human intervention in the training impossible or the results unpredictable. Therefore, common denoising neural networks are of an "end-to-end" structure. The so-called "end-to-end" structure means that the input of the denoising neural network is the image before denoising, and the output is the image after denoising.

[0003] In actual applications, there are usually specific requirements for the denoising strength, and such requirements for the denoising strength are different for different applications and different scenarios. For example, for a consumer electronics camera user, it is usually expected that the denoising strength is moderate so that the overall image looks more grainy. For a security monitoring application, it is usually expected that the denoising strength is high and there is no unnecessary noise. However, the existing neural network denoising algorithm can only change the denoising strength by retraining different networks. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a denoising neural network training method and a denoising method to solve the problem that the existing neural network denoising algorithm can only change the denoising strength by retraining different networks.

[0005] The denoising neural network training method provided by the embodiments of the present application includes n denoising units, and n is a positive integer greater than or equal to 2.

[0006] The method includes:

[0007] The image without noise is taken as a base image;

[0008] The first noise is added to the base image to obtain a first noise image;

[0009] Part of the noise is removed from the first noise to obtain a second noise, and the second noise is added to the base image to obtain a second noise image;

[0010] The first noise image is taken as a training input image, and the second noise image is taken as a training output image, and the first noise reduction unit is trained to obtain the trained first noise reduction unit.

[0011] In the technical solution, the noise reduction neural network includes n noise reduction units, each of which is trained separately. The training of the first noise reduction unit includes: taking an image without noise as a base image; adding first noise to the base image to obtain a first noise image; removing part of the noise from the first noise image to obtain a second noise image; adding second noise to the base image to obtain a second noise image; taking the first noise image as a training input image, and taking the second noise image as a training output image, and training the first noise reduction unit to obtain the trained first noise reduction unit. The second noise image still has part of the noise. Through the training method, the noise reduction strength of the noise reduction unit can be controlled. Then, by combining multiple noise reduction units, a network with different noise reduction strengths can be provided without training different networks to change the noise reduction strength.

[0012] In some optional embodiments, the first noise reduction unit includes a first convolutional layer and other convolutional layers.

[0013] The method further includes:

[0014] The other convolutional layers in the first noise reduction unit are replaced by a mapping table.

[0015] In the technical solution, the other convolutional layers in the first noise reduction unit are replaced by a mapping table, and only the first convolutional layer is retained, thereby avoiding the problem of large consumption of computing resources caused by too many convolutional layers, and greatly reducing the amount of calculation.

[0016] In some optional embodiments, the mapping table is configured to: input all values output by the first convolutional layer to the other convolutional layers after traversal, and fill the output values of the other convolutional layers into the mapping table.

[0017] In some optional embodiments, inputting all values output by the first convolutional layer to the other convolutional layers after traversal, and filling the output values of the other convolutional layers into the mapping table, includes:

[0018] Quantizing the feature map output by the first convolutional layer by fixed-point numbers, inputting all quantized values to the other convolutional layers after traversal, and filling the output values of the other convolutional layers into the mapping table.

[0019] In the technical solution, the feature map output by the first convolutional layer is quantized by fixed-point numbers, all quantized values are input to the other convolutional layers after traversal, and the output values of the other convolutional layers are filled into the mapping table. The storage size of the mapping table is reduced by quantizing the feature map by fixed-point numbers.

[0020] In some optional embodiments, the method further includes:

[0021] In the process of noise reduction by the first noise reduction unit, the actual noise image is input into the first noise reduction unit;

[0022] The first convolutional layer of the first noise reduction unit calculates a feature map for the actual noise image;

[0023] The feature map is input into the mapping table, and interpolation is performed using the values in the mapping table to obtain a noise reduction result.

[0024] In the above technical solution, after training, the first noise reduction unit is used in actual use. The actual noise image is input into the first noise reduction unit, and the first convolutional layer of the first noise reduction unit calculates a feature map for the actual noise image. Then the feature map is input into the mapping table, and interpolation is performed using the values in the mapping table. The interpolation algorithm here includes tetrahedral interpolation, etc. Finally, the noise reduction result of the first noise reduction unit is obtained.

[0025] In some optional embodiments, when the first noise reduction unit is trained, the loss function used is a Pixel-wise loss function between the output and the classification accuracy.

[0026] In the above technical solution, the Pixel-wise loss function calculates the pixel loss between the predicted image and the target image. Most of the loss functions, such as MSE or L2 loss, MAE or L1 loss, cross-entropy loss, etc., can be applied to the prediction between each pair of pixels of the target variable. Since these loss functions evaluate the class prediction of each pixel vector separately and then average all pixels, they assert that each pixel in the image has the same learning ability.

[0027] In some optional embodiments, the method further comprises:

[0028] Adding the i-th noise to the base image to obtain an i-th noise image, i = 1, 2, …, n;

[0029] Removing part of the i-th noise to obtain an i+1-th noise, and adding the i+1-th noise to the base image to obtain an i+1-th noise image;

[0030] Taking the i-th noise image as a training input image and the i+1-th noise as a training output image, the i-th noise reduction unit is trained to obtain a trained i-th noise reduction unit.

[0031] In the technical solution, when training the n denoising units respectively, the first denoising unit is required to denoise the first noise image to obtain a second noise image, the second denoising unit is required to further denoise the second noise image to obtain a third noise image, and so on, until the nth denoising unit is required to further denoise the nth noise image to obtain an (n+1)th noise image. In actual use of the denoising neural network, the first j denoising units can be combined according to the required denoising strength, and the actual noise image is input into the denoising neural network including the first j denoising units to obtain a denoising result suitable for the denoising strength.

[0032] The denoising method based on the denoising neural network provided in the embodiments of the present application comprises:

[0033] The first j trained denoising units are selected to construct a denoising neural network, and j is a positive integer less than or equal to n.

[0034] The actual noise image is input into the denoising neural network to obtain a denoised image.

[0035] In the technical solution, in actual use of the denoising neural network, the first j denoising units can be combined according to the required denoising strength, and the actual noise image is input into the denoising neural network including the first j denoising units to obtain a denoising result suitable for the denoising strength.

[0036] The electronic device provided in the embodiments of the present application comprises a processor and a memory, the memory stores machine readable instructions executable by the processor, and the machine readable instructions are executed by the processor to perform the method described in any of the above.

[0037] The computer readable storage medium provided in the embodiments of the present application stores a computer program, and the computer program is run by a processor to perform the method described in any of the above. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments of the present application. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0039] Figure 1 The training method of the denoising neural network provided in the embodiments of the present application is shown in the flowchart.

[0040] Figure 2 The method steps flowchart for denoising by using the trained first denoising unit in the embodiments of the present application is shown.

[0041] Figure 3 A denoising method based on a denoising neural network is provided for an embodiment of the present application, and a step flowchart is shown in FIG. 1.

[0042] Figure 4 A denoising neural network is provided for an embodiment of the present application, and a schematic diagram is shown in FIG. 2.

[0043] Figure 5 A structure schematic diagram of a denoising unit is provided for an embodiment of the present application, and is shown in FIG. 3.

[0044] Figure 6 A structure schematic diagram of a denoising unit is provided for another embodiment of the present application, and is shown in FIG. 4.

[0045] Figure 7 A possible structure schematic diagram of an electronic device is provided for an embodiment of the present application, and is shown in FIG. 5.

[0046] Icon: 1-processor, 2-memory, 3-communication interface, 4-communication bus. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be described below with reference to the accompanying drawings in the embodiments of the present application.

[0048] High-quality images play a key role in computer vision-related applications, such as object detection, image segmentation, scene recognition, etc. However, due to the physical characteristics of sensors and harsh shooting environments (such as insufficient ambient brightness), the images captured by cameras often have very large noise. Therefore, how to denoise the images greatly affects vision-related applications. Traditional denoising algorithms are difficult to have a good denoising effect on images in scenes with insufficient brightness (very large noise). With the popularity of convolutional neural networks (CNN), end-to-end denoising algorithms based on CNN have begun to emerge. However, there are two major problems with CNN-based denoising algorithms: 1. The network itself requires a large amount of computing resources. 2. The denoising intensity cannot be controlled through the network.

[0049] In one or more embodiments of the present application, a training method and a denoising method of a denoising neural network are provided. Compared with the problem of large consumption of computing resources caused by the introduction of a large number of convolutional layers in existing neural network denoising algorithms, one or more embodiments of the present application only need a small number of convolutional layers while retaining the original denoising effect, greatly reducing the amount of calculation. In addition, in the face of the difficulty that existing neural network denoising algorithms can only change the denoising intensity by retraining different networks, i.e., the problem that the network design itself cannot adjust the denoising intensity, the solution proposed by one or more embodiments of the present application can explicitly adjust the denoising intensity of the input image.

[0050] Please refer to Figure 1 , Figure 1A training method of a noise reduction neural network is provided in the embodiments of the present application. The noise reduction neural network comprises n noise reduction units, and n is a positive integer greater than or equal to 2. The method comprises the following steps:

[0051] In step 101, an image without noise is taken as a base image.

[0052] In step 102, first noise is added to the base image to obtain a first noise image.

[0053] In step 103, part of the first noise is removed to obtain second noise, and the second noise is added to the base image to obtain a second noise image.

[0054] In step 104, the first noise image is taken as a training input image, the second noise image is taken as a training output image, the first noise reduction unit is trained, and a trained first noise reduction unit is obtained.

[0055] In the embodiments of the present application, the noise reduction neural network comprises n noise reduction units, and each noise reduction unit is trained separately. The training of the first noise reduction unit comprises the following steps: an image without noise is taken as a base image; first noise is added to the base image to obtain a first noise image; part of the first noise is removed to obtain second noise, and the second noise is added to the base image to obtain a second noise image; the first noise image is taken as a training input image, the second noise image is taken as a training output image, the first noise reduction unit is trained, and a trained first noise reduction unit is obtained. The second noise image still has part of the noise. The training method can control the noise reduction strength of the noise reduction unit. Then, a network with different noise reduction strengths can be provided by combining multiple noise reduction units, and different networks do not need to be retrained to change the noise reduction strength.

[0056] In some optional embodiments, the first noise reduction unit comprises a first convolutional layer and other convolutional layers. The method further comprises the following step 105: replacing the other convolutional layers in the first noise reduction unit with a mapping table.

[0057] In the embodiments of the present application, the other convolutional layers in the first noise reduction unit are replaced with a mapping table, and only the first convolutional layer is retained, thereby avoiding the problem of large consumption of computing resources caused by too many convolutional layers, and greatly reducing the amount of calculation.

[0058] In some optional embodiments, the mapping table is configured to input all values output by the first convolutional layer to the other convolutional layers after traversal, and fill the output values of the other convolutional layers into the mapping table.

[0059] In some optional embodiments, all values output by the first convolutional layer are traversed and input into other convolutional layers, and the output values of the other convolutional layers are filled into the mapping table, specifically including: quantizing the feature map output by the first convolutional layer by fixed-point decimal, traversing all the quantized values and inputting them into other convolutional layers, and filling the output values of the other convolutional layers into the mapping table.

[0060] In the embodiments of the present application, the feature map output by the first convolutional layer is quantized by fixed-point decimal, all the quantized values are traversed and input into other convolutional layers, and the output values of the other convolutional layers are filled into the mapping table. The storage size of the mapping table is reduced by quantizing the feature map by fixed-point decimal.

[0061] In some optional embodiments, the loss function used by the first denoising unit during training is a Pixel-wise loss function between its output and classification accuracy.

[0062] In the embodiments of the present application, the Pixel-wise loss function calculates the loss between the pixels of the predicted image and the target image. Most of the loss functions, such as MSE or L2 loss, MAE or L1 loss, cross-entropy loss, etc., can be applied to predict between each pair of pixels of the target variable. Since these loss functions evaluate the class prediction of each pixel vector separately and then average all pixels, they assert that each pixel in the image has the same learning ability.

[0063] In some optional embodiments, the method further includes: adding the i-th noise to the base image to obtain an i-th noise image, i = 1, 2, …, n; removing part of the i-th noise to obtain an i+1-th noise, adding the i+1-th noise to the base image to obtain an i+1-th noise image; taking the i-th noise image as a training input image, taking the i+1-th noise as a training output image, and training the i-th denoising unit to obtain a trained i-th denoising unit.

[0064] In the embodiments of the present application, when training the n denoising units respectively, the first denoising unit is required to denoise the first noise image to obtain a second noise image, the second denoising unit is required to further denoise the second noise image to obtain a third noise image, and so on, until the n-th denoising unit is required to further denoise the n-th noise image to obtain an n+1-th noise image. In actual use of the denoising neural network, according to the required denoising strength, the first j denoising units can be selected for combined use, and the actual noise image is input into the denoising neural network including the first j denoising units to obtain a denoising result suitable for the denoising strength.

[0065] Please refer to Figure 2 , Figure 2 The method steps flowchart for denoising by the trained first denoising unit in the present embodiments specifically includes:

[0066] Step 201, inputting the actual noise image into the first denoising unit when denoising by using the first denoising unit;

[0067] Step 202, calculating a feature map by the first convolutional layer of the first denoising unit on the actual noise image;

[0068] Step 203, inputting the feature map into the mapping table and interpolating by using the values in the mapping table to obtain a denoising result.

[0069] In the embodiment of the present application, when the trained first denoising unit is actually used, the actual noise image is input into the first denoising unit, the feature map is calculated by the first convolutional layer of the first denoising unit on the actual noise image, and then the feature map is input into the mapping table and interpolated by using the values in the mapping table, wherein the interpolation algorithm includes tetrahedron interpolation and the like, and finally the denoising result of the first denoising unit is obtained.

[0070] Please refer to Figure 3 , Figure 3 A denoising method based on a denoising neural network provided in the embodiment of the present application is shown in a flow chart, which specifically includes:

[0071] Step 301, selecting the first j trained denoising units to construct a denoising neural network, j being a positive integer less than or equal to n;

[0072] Step 302, inputting the actual noise image into the denoising neural network to obtain a denoised image.

[0073] In the embodiment of the present application, when the denoising neural network is actually used, the first j denoising units can be combined for use according to the required denoising strength, the actual noise image is input into the denoising neural network including the first j denoising units, and the denoising result suitable for the denoising strength is obtained.

[0074] Please refer to Figure 4 , Figure 4 A denoising neural network provided in the embodiment of the present application is shown in a schematic diagram. The denoising units in the denoising neural network please refer to Figure 5 , Figure 5A structure diagram of a noise reduction unit provided by an embodiment of the present application is shown in FIG. 1. The algorithm provided by the embodiment is a segmented design idea. Taking single-channel image noise reduction as an example, the input of the algorithm is a noise image in HxW format, and the output of each denoising block is a noise-reduced image in HxW format. In each denoising block, there are 5 layers of convolutional networks, and the convolution of each layer uses relu as an activation function and stride is 1. The parameters of the convolutional networks are as follows: 5x5 convolution kernel, 1 input channel, and 3 output channels; 3x3 convolution kernel, 3 input channels, and 64 output channels; 3x3 convolution kernel, 64 input channels, and 64 output channels; 3x3 convolution kernel, 64 input channels, and 64 output channels; and 3x3 convolution kernel, 64 input channels, and 1 output channel.

[0075] To control the noise reduction strength, the noise reduction strength of each denoising block is different. In the entire network, each denoising block has 5 layers of convolutional networks, and if there are n blocks, there will be 5x n layers of convolutional networks. By controlling the number of denoising blocks through which the input noise image is processed, the noise reduction strength of the neural network algorithm on the image can be explicitly controlled.

[0076] During network training, images with different noise reduction strengths can be used as the ground truth (classification accuracy of the training set for supervised learning technology) corresponding to different denoising blocks. Specifically, a noise-free (or a small amount of noise) image in the data set is used as the base, a first noise image is obtained by adding noise, the second noise image output by the first denoising block has the largest residual noise, the third noise image has less residual noise than the second noise image, and so on, until the n-th denoising block outputs the n+1-th noise image with the smallest residual noise. During network training, each denoising block can be trained independently, and the L2 pixel-wise loss between the network output and the ground truth is used for training.

[0077] These denoising blocks can also be used independently or in combination. For example, the first j denoising blocks can be combined to form a denoising neural network, the input image of the denoising neural network is first input into the first denoising block for noise reduction, the output image of the first denoising block is further input into the second denoising block for noise reduction, the output image of the second denoising block is further input into the third denoising block for noise reduction, and so on, until the output image of the i-1-th denoising block is input into the i-th denoising block for the last step of noise reduction, and finally the noise reduction result is obtained.

[0078] Please refer to Figure 6 , Figure 6 The structure diagram of the denoising unit provided by another embodiment of the present application is shown. In order to significantly reduce the calculation amount of the algorithm, in specific deployment, a mapping table LUT and a single-layer convolution are combined to replace the denoising unit denoising block as shown in Figure 6 The coefficients of the first 5x5 convolution layer in the denoising unit denoising block are the first-layer convolution trained in specific deployment. The feature map obtained after the convolution is of size HxWx3. In order to reduce the storage size of the mapping table LUT, the feature map can be quantized as a fixed-point decimal. Taking the input input image as 8 bits and the feature map after one layer of convolution as 8 bits as an example, the data amount stored in the mapping table LUT is 2^8x2^8^2^8^8bit=16MB. The mapping table LUT is used to complete the convolution calculation of the last four layers in the denoising unit denoising block, and the value of the corresponding position in the mapping table LUT is the output of the trained network. After traversing all the values, the output values of the last four layers of the convolution network are filled into the mapping table LUT. In this embodiment, the mapping table LUT is used after the 5x5 convolution layer to improve the receptive field. In actual use, the result calculated based on the 5x5 convolution layer is used to obtain the final output result by interpolation using the values in the mapping table LUT. The interpolation method can be tetrahedral interpolation and the like.

[0079] In summary, the present embodiment innovatively proposes a denoising algorithm combining a neural network and a mapping table, which significantly reduces the calculation amount of the existing neural network denoising scheme. In addition, a segmented neural network structure is proposed to explicitly control the denoising strength, which can further reduce the calculation amount when the denoising strength requirement is small.

[0080] Figure 7 A possible structure of an electronic device provided by an embodiment of the present application is shown. Referring to Figure 7 , the electronic device includes a processor 1, a memory 2 and a communication interface 3, which are interconnected and communicate with each other through a communication bus 4 and / or other forms of connection mechanism (not shown).

[0081] The memory 2 includes one or more (only one is shown in the figure), which can be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc. The processor 1 and other possible components can access, read and / or write data in the memory 2.

[0082] The processor 1 includes one or more (only one is shown in the figure), which can be an integrated circuit chip with signal processing capability. The processor 1 described above can be a general-purpose processor, including a central processing unit (CPU), a micro controller unit (MCU), a network processor (NP) or other conventional processors; it can also be a special-purpose processor, including a neural network processing unit (NPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. Moreover, when the processor 1 is multiple, part of them can be general-purpose processors and the other part can be special-purpose processors.

[0083] The communication interface 3 includes one or more (only one is shown in the figure), which can be used for direct or indirect communication with other devices to interact with data. The communication interface 3 can include interfaces for wired and / or wireless communication.

[0084] The one or more computer program instructions can be stored in the memory 2, and the processor 1 can read and run the computer program instructions to implement the method provided by the embodiments of the present application.

[0085] It can be understood that, Figure 7 The structure shown is only schematic, and the electronic device can further include more or fewer components than those shown in the figures, or have a different configuration of components than those shown in the figures. Figure 7 The structure shown is only schematic, and the electronic device can further include more or fewer components than those shown in the figures, or have a different configuration of components than those shown in the figures. Figure 7 The structure shown is only schematic, and the electronic device can further include more or fewer components than those shown in the figures, or have a different configuration of components than those shown in the figures. Figure 7 The components shown in the figures can be implemented in hardware, software, or a combination thereof. The electronic device can be a physical device, such as a PC, a notebook computer, a tablet computer, a mobile phone, a server, an embedded device, etc., or a virtual device, such as a virtual machine, a virtualization container, etc. Moreover, the electronic device is not limited to a single device, but can also be a combination of multiple devices or a cluster of a large number of devices.

[0086] The embodiments of the present application also provide a computer readable storage medium, which stores computer program instructions, and the computer program instructions are read and run by a processor of a computer to execute the method provided by the embodiments of the present application. For example, the computer readable storage medium can be implemented as Figure 7 The memory 2 in the electronic device in the figures.

[0087] In the embodiments provided by the present application, it should be understood that the disclosed device and method can be implemented by other ways. The device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and there can be another division way in actual implementation, and some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed mutual elements can be indirect coupling or communication connection through some communication interfaces, devices or units, and can be electrical, mechanical or other forms.

[0088] In addition, the units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e. they can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments of the present application.

[0089] Furthermore, the functional modules in each of the embodiments of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0090] In this document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions.

[0091] The above description is merely illustrative of the application and not in limitation of the principles of the application. Numerous modifications and adaptations thereof will be readily apparent to those skilled in the art without departing from the spirit and scope of the application as defined in the following claims.

Claims

1. A method for training a noise reduction neural network, the method comprising: The noise reduction neural network comprises n noise reduction units, n being a positive integer greater than or equal to 2; The method comprises: Taking an image without noise as a base image; Adding first noise to the base image to obtain a first noise image; Removing part of the first noise to obtain second noise, and adding the second noise to the base image to obtain a second noise image; Training the first noise reduction unit by taking the first noise image as a training input image and the second noise image as a training output image, to obtain a trained first noise reduction unit; The method further comprises: Adding ith noise to the base image to obtain an ith noise image, i = 1, 2, …, n; Removing part of the ith noise to obtain (i+1)th noise, and adding the (i+1)th noise to the base image to obtain an (i+1)th noise image; Training the ith noise reduction unit by taking the ith noise image as a training input image and the (i+1)th noise as a training output image, to obtain a trained ith noise reduction unit; Further comprising: Selecting the first j trained noise reduction units to construct a noise reduction neural network, j being a positive integer less than or equal to n; Inputting an actual noise image into the noise reduction neural network to obtain a noise reduction image.

2. The method of claim 1, wherein, The first noise reduction unit comprises a first convolutional layer and other convolutional layers; The method further comprises: Replacing the other convolutional layers in the first noise reduction unit with a mapping table.

3. The method of claim 2, wherein, Wherein, The mapping table is configured to input all values output by the first convolutional layer into the other convolutional layers after traversal, and fill the output values of the other convolutional layers into the mapping table.

4. The method of claim 3, wherein, The inputting all values output by the first convolutional layer into the other convolutional layers after traversal, and filling the output values of the other convolutional layers into the mapping table, comprises: Quantizing the feature maps output by the first convolutional layer by fixed-point numbers, inputting all quantized values into the other convolutional layers after traversal, and filling the output values of the other convolutional layers into the mapping table.

5. The method of claim 4, wherein, Further comprising: When using the trained first noise reduction unit for noise reduction, inputting an actual noise image into the trained first noise reduction unit; The first convolutional layer of the trained first noise reduction unit calculates feature maps for the actual noise image; Inputting the feature maps into the mapping table of the trained first noise reduction unit, and using the values in the mapping table for interpolation to obtain a noise reduction result.

6. The method of claim 1, wherein, When training the first noise reduction unit, the loss function used is a Pixel-wise loss function between its output and classification accuracy.

7. An electronic device, comprising: Comprise: A processor and a memory, the memory stores machine readable instructions executable by the processor, and the machine readable instructions are executed by the processor to perform the method of any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, and the computer program is executed by the processor to perform the method of any one of claims 1-6.

Images