An Image Denoising Method Based on a Lightweight Double Convolutional Neural Network

Abstract

This solution discloses an image denoising method based on a lightweight double convolutional neural network in the field of intelligent image processing technology. A noise mask convolutional neural network (MCNN) and a denoising convolutional neural network (DCNN) are respectively built. MCNN aims to achieve accurate noise point marking, while DCNN aims to achieve high-performance denoising. The "noisy image-noise mask" pair dataset and the "noisy image-clean image" pair dataset are used to train MCNN and DCNN. The experimental results show that the misjudgment rates marked by MCNN are reduced by 77.79%, 77.75% and 30.60% compared with the extreme point marking method, the mean value marking method and the extreme value image block marking method respectively. The peak signal-to-noise ratio of the denoised image is improved by 4.84% compared with the traditional method, and the information loss is reduced by 17.89%. In addition, the operation complexity of the network is reduced by 3.98 times compared with the traditional CNN, and the performance of salt-and-pepper denoising is also improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent image processing, and particularly relates to an image denoising method based on a lightweight double convolutional neural network. Background Art

[0002] With the advent of the intelligent era, the application of image processing has gradually penetrated into emerging fields such as target tracking, license plate recognition, face recognition, driverless, and intelligent medical imaging. Images are easily contaminated by salt-and-pepper noise during acquisition, storage, and transmission, and need to be pre-processed before further specific applications. The research on salt-and-pepper denoising has a long history. Starting from median filtering in the 1980s, to various pixel gray-level based filtering methods in the first decade of this century, to the artificial neural network denoising methods that have emerged in recent years, the denoising performance has been gradually improved. Median filtering replaces the gray level of the noise point with the median of all pixel gray levels in the sliding window. This method has two main disadvantages. One is that the indiscriminate processing of all gray levels causes serious information loss. The other is that noise residues will occur when the noise density is slightly high. Although this method does not achieve the ideal denoising effect, the exploration around it provides a feasible idea for the improvement of subsequent denoising methods.

[0003] The first idea is to improve the accuracy of noise point marking. In order to avoid indiscriminate processing, it is necessary to study noise point marking, aiming to only process the pixels marked as noise points and directly output the normal pixels, so that the information loss can be reduced. For this purpose, the extreme point marking method, the mean marking method, and the extreme value image block marking method have been proposed in the industry. The extreme point marking method marks the extreme points in the noise image. For an 8-bit image, the pixels with gray levels of 0 or 255 are marked as noise points. Generally, the marking of this method is accurate, but if there are many extreme pixels in the image itself, this method will misjudge these extreme pixels as noise points. The mean marking method expands the field of view from a point to a window pixel array centered on the scanning point, and discriminates the noise points by calculating the average gray level of the pixels in the window. Although this method improves the marking accuracy, it is prone to failure when encountering a large area of extreme pixel image regions. The extreme value image block marking method discriminates the noise points by comparing the number of normal pixels and the number of extreme pixels in the window, and improves the marking accuracy at low noise density. However, in the case of extremely high noise density, there are many noise points everywhere, and the size relationship of the number of noise points no longer has distinguishable features. Therefore, this method is prone to failure at high density. Therefore, in the noise point marking step, in-depth research is needed to improve the accuracy and stability of noise point marking.

[0004] The second approach is to improve the method of updating noisy grayscale levels so that noise can be completely removed even at high densities. The literature has proposed median filtering with all normal pixel units. This method does reduce information loss, but there is still significant room for improvement in the visual effect of the image. In addition to median filtering and its derivative methods, the industry has also studied weighted filters, which use the sum of the weights of the grayscale levels of adjacent pixel units to replace the grayscale level of the noisy pixel unit. This method improves the visual effect of the processed image, but residual noise will occur when the noise density is too high. In addition, the industry has proposed using probability filtering for salt-and-pepper noise removal, which uses the grayscale level of the adjacent pixel with the highest probability to replace the noisy grayscale level. However, this method is prone to failure when dealing with high-density noise. To emphasize the influence of pixels at key positions, the industry has proposed iterative filtering, which repeatedly uses the pixel unit closest to the noise to generate a new grayscale level. However, since the nearest neighbor pixel may be noisy, the processing effect of this method is still not ideal.

[0005] In recent years, with the rapid advancement of the computing power of graphics processing units (GPUs) and the convenience provided by the Internet for obtaining large amounts of image data, research on artificial neural networks has continued to make progress. However, no salt-and-pepper noise removal method applying a dual convolutional neural network has been proposed based on the above two ideas combined with artificial neural network technology. Summary of the Invention

[0006] In order to improve the accuracy of salt-and-pepper noise marking, the present invention proposes a salt-and-pepper noise marking method based on a lightweight dual convolutional neural network.

[0007] An image denoising method based on a lightweight dual convolutional neural network in this solution includes:

[0008] S1. Convert the color image into a black-and-white image, and then intercept a number of image patches;

[0009] S2. Add salt-and-pepper noise with a random density to the intercepted image patches to obtain a noisy image; distinguish and mark the positions with and without salt-and-pepper noise added respectively to generate a noise mask;

[0010] S3. (1) Combine the noisy image and the noise mask to form a corresponding "noisy image-noise mask" pair, which is used as the dataset for training the MCNN to obtain the MCNN network model;

[0011] (2) Combine the noisy image and the clean image to form a corresponding "noisy image-clean image" pair, which is used as the dataset for training the DCNN to obtain the DCNN network model;

[0012] S4. Denoise the noisy image: Use MCNN to label the noise points in the image. For the pixels labeled as noise points by MCNN, use the DCNN network model to denoise the noisy image.

[0013] The clean image described in this solution is the image block without added salt-and-pepper noise after being intercepted in S1.

[0014] Furthermore, the MCNN network model includes several convolutional layers, preferably in the range of 8 to 12 convolutional layers, and optimally 10 convolutional layers. The first and the last convolutional layers are conventional convolutional layers. The first convolutional layer generates several convolutional tensors from the input noisy image, preferably 50 to 70 convolutional tensors, and optimally 64 convolutional tensors. The last convolutional layer generates 1 noise mask for all the input tensors. Between the two conventional convolutional layers are multiple repeated depthwise separable convolutional layers, preferably 5 to 10 repeated depthwise separable convolutional layers, and optimally 8 repeated depthwise separable convolutional layers.

[0015] Furthermore, the DCNN network model includes several convolutional layers, preferably in the range of 15 to 20 convolutional layers, and optimally 17 convolutional layers. The first convolutional layer and the last convolutional layer are conventional convolutional layers. The first convolutional layer generates several convolutional tensors from the input noisy image, preferably 48 to 64 convolutional tensors, and optimally 64 convolutional tensors. The last convolutional layer generates 1 noise mask for all the input tensors. Between the two conventional convolutional layers are multiple repeated depthwise separable convolutional layers, preferably 12 to 18 repeated depthwise separable convolutional layers, and optimally 15 repeated depthwise separable convolutional layers.

[0016] In the MCNN network model and the DCNN network model, the first conventional convolutional layer and all depthwise separable convolutional layers perform ReLU activation.

[0017] Furthermore, the training cycle of the MCNN and DCNN is 50 training cycles.

[0018] Furthermore, the training duration of the MCNN is 74 minutes, and the training duration of the DCNN is 146 minutes.

[0019] The beneficial technical effects of the present invention are as follows: The innovation of the present invention based on the lightweight double convolutional neural network lies in: (1) Using one convolutional neural network each in the noise point marking link and the noise removal link, realizing accurate noise point marking and high-performance denoising; (2) Using depthwise separable convolution instead of conventional convolution in the middle layers of the two convolutional neural networks, greatly reducing the computational complexity.

[0020] The experimental results show that the misjudgment rates of the MCNN markers are respectively reduced by 77.79%, 77.75% and 30.60% compared with the pole marking method, the mean marking method and the extreme value image block marking method. The peak signal-to-noise ratio of the denoised image is increased by 4.84% compared with the traditional method, and the information loss is reduced by 17.89%. In addition, the operation complexity of the network is reduced by 3.98 times compared with the traditional CNN. This method improves the performance of salt-and-pepper denoising while reducing the operation complexity of the network. Description of the Drawings

[0021] Figure 1 Method for constructing a data set;

[0022] Figure 2 MCNN model;

[0023] Figure 3 DCNN model;

[0024] Figure 4 Depthwise separable convolution;

[0025] Figure 5 Schematic diagram of the image block intercepting method;

[0026] Figure 6 Pair of "noisy image - noise mask" randomly intercepted;

[0027] Figure 7 Pair of "noisy image - clean image" randomly intercepted;

[0028] Figure 8 Comparison of noisy point marked images;

[0029] Figure 9 Comparison of denoising effects of different algorithms under the same noisy point marking. Detailed Implementation Manner

[0030] In order to improve the accuracy of salt-and-pepper noise point marking, the present invention proposes a salt-and-pepper noise point marking method based on a lightweight double convolutional neural network. The content of the present invention includes:

[0031] ■ Propose a method for constructing a data set

[0032] ■ Propose a lightweight convolutional neural network for noise point marking

[0033] ■ Propose a lightweight convolutional neural network for noise removal

[0034] (1) Propose a method for constructing a data set

[0035] Based on the above image block interception and generation of noise masks, the method for constructing a data set is as follows Figure 1As shown in the figure, the 91image dataset is selected as the original data, and the color images are converted to black and white and the resolution is adjusted. 25 image patches are cropped from each image at a step size of 20, and salt-and-pepper noise with random density is added to these image patches to obtain noisy image patches. At the same time, the positions where the noise is added are marked as white points, and the other positions are marked as black points to generate a noise mask corresponding to the noisy image. Then, the noisy image patches and the noise mask patches are combined together to form a corresponding "noisy image-noise mask" pair. Finally, the noisy image patches and the clean image patches are combined together to form a corresponding "noisy image-clean image" pair. Among them, the "noisy image-noise mask" pair will be used as the dataset for training the MCNN, and the "noisy image-clean image" pair will be used as the dataset for training the DCNN.

[0036] (2) A salt-and-pepper noise marking method based on a lightweight convolutional neural network is proposed

[0037] Figure 2 The MCNN model adopted in this paper is used to train the noise mask. This network uses the "noisy image-noise mask" pair dataset as the training data, and the mean square error (mse) between the network output image and the ideal noise mask is used as the loss function. The network model includes 10 convolutional layers, where the first convolutional layer and the last convolutional layer are both conventional convolutional layers. The first convolutional layer generates 64 convolutional tensors from the input noisy image, and the last convolutional layer generates 1 noise mask from the input 64 tensors. Between the head and the tail are 8 repeated depthwise separable convolutional layers.

[0038] Figure 3 The DCNN model adopted in this paper uses the "noisy image-clean image" pair dataset as the training data, and the mse between the network output image and the clean image is used as the loss function. The network model includes 17 convolutional layers, where the first layer is a conventional convolution that generates 64 convolutional tensors from the input noisy image. The last convolutional layer generates 1 noise mask from the input 64 tensors. Between the head and the tail are 15 repeated depthwise separable convolutional layers.

[0039] MobileNet-v1 has proven that the complexity of depthwise separable convolution is reduced by 8-9 times compared to conventional convolution. As Figure 4 shown, first, a 3x3 convolutional kernel is used to process 64 input channels, and then, batch normalization and ReLU activation are performed. Then, a 1x1 convolutional kernel is used to process all channels in turn to obtain 64 processed channels, and this step is called point convolution. Finally, 64 channels are output after batch normalization and activation.

[0040] To demonstrate the specific implementation method of the present invention, specific implementation manners are provided. It should be noted that the data used in the following steps will be specified, and relevant data modifications may not greatly affect the implementation effect and still fall within the scope of protection of this patent. The steps sequentially executed by the present invention include: Step 1, namely dataset preparation, including generating "noisy image-noise mask" pairs and generating "noisy image-clean image" pairs; Step 2, namely machine training, including training of MCNN and DCNN; Step 3, namely denoising the input noisy image.

[0041] Step 1: Dataset preparation

[0042] (1) Image patch extraction

[0043] The method of extracting image patches is as Figure 5 shown. In a clean image, starting from the upper left corner of the image, image patches with a resolution of 70*70 are extracted at a step size of 20. For an image with a resolution of 200*200, it is extracted 5 times horizontally and vertically, scanned from left to right and from top to bottom, and 25 image patches of 70*70 are obtained. In order to be able to completely characterize the image features, the resolution of the image patches should not be too small. In order to avoid too large a number of training parameters, the resolution of the image patches should not be too large. Therefore, an image patch resolution of 70*70 is selected.

[0044] (2) Generating "noisy image-noise mask" pairs

[0045] In order to train the noise mask, "noisy image-noise mask" needs to be fed to MCNN. The network will process the noisy image therein, output the noise mask, compare the output with the fed-in noise mask, and generate a loss function. Based on the above-generated 70*70 image patches, salt-and-pepper noise with a random density is added to generate a noisy image. At the same time, the position where noise is added each time is marked as 1, and the pixels in the image where no noise is added are marked as 0, thus generating a noise mask. It should be noted that the noisy image and the noise mask correspond one by one to form a "noisy image-noise mask" pair. In addition, in order to enhance the robustness of the dataset, the density of noise added to each image patch is random, and the noise density value range is a random number between 0.1 and 0.9. Figure 6 is a randomly intercepted "noisy image-noise mask" pair. It can be seen from it that the noisy image is obtained by adding noise to the image patch, and the noise mask marks the positions where noise is added as white dots and the normal pixels as black dots.

[0046] (3) Generating "noisy image-clean image" pairs

[0047] To achieve high-performance denoising, it is necessary to train a DCNN, and the data fed to the DCNN is a "noisy image - clean image" pair. Among them, the noisy image is the input of the DCNN, and the output of this convolutional neural network will be compared with the clean image to generate a loss function. Figure 7 They are randomly cropped "noisy image - clean image" pairs. Among them, the noisy image is obtained by adding salt-and-pepper noise with a random density to the clean image, and the value range of the noise density is 0.1 - 0.9. The resolution of these image pairs is 70 * 70, the same as the noise mask.

[0048] Step 2: Machine training

[0049] Using an RTX 3080 GPU, the keras component is called on the TensorFlow platform to build MCNN and DCNN respectively. The loss function used is mse, and the initial value of the learning rate is set to 0.001. If the value of the loss function does not decrease after multiple epochs, learning rate decay is performed, and the decay factor is 0.2. The respective datasets are imported, and MCNN and DCNN are trained for 50 epochs. The training of MCNN takes 74 minutes, and the training of DCNN takes 146 minutes, as shown in Table 1. Since MCNN has 7 fewer convolutional layers than DCNN, the training time required is less.

[0050] Table 1 Training of convolutional neural networks

[0051]

[0052] Step 3: Denoise the input noisy image

[0053] (1) Noise point marking

[0054] Input the noisy image, and use MCNN to mark the noise points in the image.

[0055] (2) Noise removal

[0056] Input the noisy image. For the pixels marked as noise points by MCNN, use DCNN for denoising. For the pixels marked as normal points by MCNN, no processing is performed.

[0057] Implementation effect

[0058] (1) Noise point marking

[0059] Three images in BSD300image are selected as experimental images in this paper. Figure 8 (a1), Figure 8 (b1) and Figure 8 (c1) The proportions of pole pixels are 13%, 10%, and 10% respectively. Add salt-and-pepper noise with densities of 0.2, 0.5, and 0.8 to these three pictures respectively, as Figure 8(a2), Figure 8 (b2) and Figure 8 (c2). In the marked image, black dots represent noise points and white dots represent normal points. The extreme point marking method marks the extreme pixel points in the noise image as noise points, so it is easy to misjudge the normal extreme points that already exist in the original image as noise points. As shown in Figure 8 (a3) and Figure 8 (b3), there are multiple black image blocks in these two marked images, and these image blocks are not noise points but highlight areas in the original image. A normal noise-marked image should be evenly distributed black dots and should not be affected by the extreme pixel points in the original image. The mean marking method judges noise by comparing the difference between the average value of pixel units in the window and the extreme value. When there are many extreme points in the original image, this comparison and judgment mechanism is likely to fail. Therefore, the processing result of the mean marking method is similar to that of the extreme point marking method. As shown in Figure 8 (a5), Figure 8 (b5) and Figure 8 (c5). In contrast, the processing result of the extreme image block marking method is better. It judges noise by the size relationship between the number of normal pixels and extreme pixels in the window, and achieves a marking effect that surpasses the previous two methods at low density. As shown in Figure 8 (a4) and Figure 8 (b4), the marked noise is evenly distributed and is hardly misled by the normal extreme points in the original image anymore. However, at high noise density, this method marks noise points as normal points, increasing the misjudgment rate. As shown in Figure 8 (c4). Compared with traditional methods, the MCNN marking method proposed in this paper extracts the features of noise points comprehensively through a convolutional neural network, discovers the internal law of noise distribution from a large number of image blocks, and realizes a more objective noise marking result through machine intelligence. As shown in Figure 8 (a6) and Figure 8 (b6), the noise is evenly distributed and there is no obvious boundary feature as in the extreme image block marking. Such a marking is more in line with the properties of uniform distribution and random distribution of noise. In addition, at high density, the marked image of this method does not have as many white dots as the marked image of the extreme image block marking, and still has a low misjudgment rate.

[0060] Table 2 Comparison of misjudgment rates of different marking methods

[0061] Noise density Pole marker Extreme value image block marker Mean marker MCNN marker 0.2 0.5525 0.0285 0.5525 0.0832 0.5 0.1089 0.0495 0.1089 0.0471 0.8 0.0248 0.1417 0.0237 0.0222 Average value 0.2287 0.0732 0.2283 0.0508

[0062] Table 2 shows the comparison of the misjudgment rates of different marking methods. The misjudgment rate (MR) is the ratio of the number of mislabeled noise points to the total number of noise points, and the lower the MR, the better. The data in Table 2 is consistent with the above analysis. The misjudgment rates of the pole marking method and the mean marking method are comparable. The misjudgment rate of the extreme value image block is relatively low, while the misjudgment rate of the MCNN marking method is the lowest. The data shows that the misjudgment rate of the MCNN marking method is reduced by 75.47%, 75.43% and 23.36% compared with the pole marking method, the mean marking method and the extreme value image block marking method respectively.

[0063] (2) Noise removal

[0064] To compare the performance of DCNN and traditional denoising methods, for the images obtained by processing all denoising methods, according to the results of MCNN noise point marking, the normal pixels are replaced by the pixels at the corresponding positions in the clean image, and the noise pixels retain the original gray levels generated by their respective denoising methods. After ablation according to this method, the visual effects of the images and the values of peak signal to noise ratio (PSNR) and mean absolute error (MAE) are compared. Figure 9 Figure shows the comparison of the visual effects of the images obtained by different methods under different noise densities. The input images are noise images with densities of 0.2, 0.5 and 0.8 respectively, covering low, medium and high levels of noise density. It can be seen that there are still noise residues in the median filtering denoised images at low densities, such as Figure 9 (a2). At high densities, the noise residue phenomenon is more serious, such as Figure 9 (b2) and Figure 9 (c2). This is because this method uses the median value of the pixel gray levels in the sliding window to replace the noise points. At a certain noise density, the median point may also be a noise point, and the probability of this situation increases as the noise density increases. The three-weight factor algorithm uses the sum of the weights of multiple pixel gray levels in an adaptive window to replace the noise points. The denoising effect is acceptable at low densities, but at high densities, due to the large misjudgment rate of the pole marking used in this method when there are many normal extreme points in the input image itself, the normal points searched may be noise points, and using noise points to replace noise points will naturally result in noise residues, such as Figure 9 (a3), Figure 9 (b3) and Figure 9 (c3). The adaptive probability filtering uses the gray level of the adjacent pixel with the highest probability to replace the noise gray level, but in the case of extremely high noise density, the adjacent pixels may be noise points, so the probability filtering is prone to failure when processing high-density noise images, such as Figure 9(c4). Among traditional non-CNN denoising methods, mean label repeated filtering performs the best. The main reason is that the misjudgment rate of mean noise labeling adopted by this method is relatively low. However, the information loss of this method still needs to be further reduced. Liang CNN performs thorough denoising through median filtering and then restores the normal image through the learning of the residual network. The noise in the image obtained by Liang CNN denoising is completely eliminated. However, due to the extensive use of median filtering operators in the first half of the network, a large amount of image details are lost, and the image processed by this method appears blurred, such as Figure 9 (a7), Figure 9 (b7) and Figure 9 (c7). Xing CNN removes noise through 17 convolutional layers. In comparison, the information loss is reduced. However, this network changes all pixel units using convolutional neural networks, leaving room for further improvement, as shown in Table 3. In addition, Chen CNN proposes to directly output normal points according to the results of pole labeling, and use the gray scale processed by convolutional neural networks for noisy points. The universality of this method is poor. Especially when there are many normal extreme points in the original image, convolutional neural network data is still used to update these pixels that do not need to be changed originally. Therefore, the information loss of this method is large, as shown in Table 3. To achieve high-performance denoising while reducing information loss, this paper uses dual CNN to improve the robustness of salt-and-pepper denoising. The task of MCNN is to achieve accurate noise labeling, and the task of DCNN is to achieve high-performance noise removal. Through the combination of two CNNs, targeted high-performance processing of noise is achieved, thus achieving the best denoising effect, such as Figure 9 (a10), Figure 9 (b10) and Figure 9 (c10).

[0065] The data in Table 3 is consistent with the above analysis. It can be seen that the denoising method proposed in this paper has obvious improvement compared with non-CNN methods and CNN methods. The peak signal-to-noise ratio (PSNR) evaluates the visual effect of the image, and the larger the better. The mean absolute error (MAE) of information loss evaluates the information loss, and the smaller the better. It can be seen from Table 3 that mean label repeated filtering is relatively good among traditional non-CNN methods, and Xing CNN and Chen CNN have achieved relatively good denoising effects among traditional CNN methods. The PSNR of the dual convolutional neural network method has increased by 8.61% and 6.04% compared with the mean label repeated filtering and Chen CNN methods respectively, and the MAE has decreased by 15.60% and 17.86% compared with the mean label repeated filtering and Chen CNN methods respectively.

[0066] Comparison of performance indicators of various denoising methods under the same noise labeling in Table 3

[0067]

Claims

1. An image denoising method based on a lightweight double convolutional neural network, characterized in that Including: S1. Convert a color image into a black-and-white image, and then intercept a number of image patches. S2. Add salt-and-pepper noise with random density to the intercepted image patches to obtain a noisy image; respectively mark the positions with and without salt-and-pepper noise added to generate a noise mask. S3. (1) Combine the noisy image and the noise mask together to form a one-to-one corresponding "noisy image-noise mask" pair, which is used as the dataset for training MCNN to obtain the MCNN network model. (2) Combine the noisy image and the clean image together to form a one-to-one corresponding "noisy image-clean image" pair, which is used as the dataset for training DCNN to obtain the DCNN network model. S4. Denoise the noisy image: Use MCNN to mark the noise points in the image. For the pixels marked as noise points by MCNN, use the DCNN network model to denoise the noisy image.

2. The image denoising method based on a lightweight double convolutional neural network according to claim 1, characterized in that: The MCNN network model includes several convolutional layers, where the first and the last convolutional layers are conventional convolutional layers. The first convolutional layer generates several convolutional tensors from the input noisy image, and the last convolutional layer generates 1 noise mask for all the input tensors. Between the two conventional convolutional layers are multiple repeated depthwise separable convolutional layers.

3. The image denoising method based on a lightweight double convolutional neural network according to claim 2, characterized in that: The MCNN network model includes 8 to 12 convolutional layers, where the first and the last convolutional layers are conventional convolutional layers. The first convolutional layer generates 50 to 70 convolutional tensors from the input noisy image, and the last convolutional layer generates 1 noise mask for all the input tensors. Between the two conventional convolutional layers are 5 to 10 repeated depthwise separable convolutional layers.

4. The image denoising method based on a lightweight double convolutional neural network according to claim 3, characterized in that: The MCNN network model includes 10 convolutional layers, where the first and the last convolutional layers are conventional convolutional layers. The first convolutional layer generates 64 convolutional tensors from the input noisy image, and the last convolutional layer generates 1 noise mask for the 64 input tensors. Between the two conventional convolutional layers are 8 repeated depthwise separable convolutional layers.

5. A method for image denoising based on a lightweight double convolutional neural network according to any one of claims 1 to 4, characterized in that: The DCNN network model includes several convolutional layers, where the first convolutional layer and the last convolutional layer are conventional convolutional layers. The first convolutional layer generates several convolutional tensors from the input noisy image, and the last convolutional layer generates 1 noise mask for all the input tensors. Between the two conventional convolutional layers are multiple repeated depthwise separable convolutional layers.

6. The image denoising method based on a lightweight double convolutional neural network according to claim 5, characterized in that: The DCNN network model includes 15 to 20 convolutional layers, where the first convolutional layer and the last convolutional layer are conventional convolutional layers. The first convolutional layer generates 48 to 64 convolutional tensors from the input noisy image, and the last convolutional layer generates 1 noise mask for all the input tensors. Between the two conventional convolutional layers are 12 to 18 repeated depthwise separable convolutional layers.

7. A method for image denoising based on a lightweight double convolutional neural network according to claim 6, characterized in that: The DCNN network model includes 17 convolutional layers, where the first convolutional layer and the last convolutional layer are conventional convolutional layers. The first convolutional layer generates 64 convolutional tensors from the input noisy image, and the last convolutional layer generates 1 noise mask for the 64 input tensors. Between the two conventional convolutional layers are 15 repeated depthwise separable convolutional layers.

8. A method for image denoising based on a lightweight double convolutional neural network according to claim 1, characterized in that: The training cycles of the MCNN and DCNN are 50 training cycles.

9. The image denoising method based on a lightweight double convolutional neural network according to claim 8, characterized in that: The training duration of the MCNN is 74 minutes, and the training duration of the DCNN is 146 minutes.

Images