A nuclear irradiation image denoising system based on supervised learning

Abstract

The application discloses a nuclear irradiation image denoising system based on supervised learning, which comprises a nuclear irradiation noise identification module and an image denoising module, both of which are realized through a deep learning network and the deep learning network for image denoising is trained through a supervised learning method; the deep learning network for realizing nuclear irradiation noise identification inputs a nuclear irradiation noise image and outputs a mask image corresponding to the nuclear irradiation noise image. In the application, the training method of the nuclear irradiation noise denoising model using the supervised learning mode is improved, the problem that nuclear irradiation noise and noise-free images are difficult to obtain in the data set when the supervised learning is used is solved, the denoising problem of the nuclear irradiation noise with complex features is converted into the denoising problem of the color block with simple features through the feature recognition and accurate positioning of the nuclear irradiation noise, and a large number of noise and noise-free paired images can be artificially and quickly generated for model training when the denoising network model is trained.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image denoising technology, and in particular to a nuclear irradiation image denoising system based on supervised learning. BACKGROUND

[0002] There are many noises in images taken by cameras, such as additive white Gaussian noise (AWGN), impulse noise (salt and pepper noise), quantization noise, Poisson noise, and speckle noise. In a nuclear radiation environment, various nuclear radiation such as alpha, beta, gamma, and neutron rays will produce various nuclear radiation noises on the images collected by the camera. Due to the large difference in formation mechanism, nuclear irradiation noise and common image noise are quite different in noise image performance.

[0003] Current industry feature recognition of nuclear irradiation noise is still based on personal experience and traditional statistical methods, and manual feature extraction and noise model design, which has great limitations in application. The disadvantages are as follows:

[0004] (1) The traditional image denoising method is manually designed for different noise types and corresponding denoising algorithms or image filters. Since noise appears randomly in images and is not fixed in position, the whole image needs to be denoised or filtered. Due to the limitations of manually designed noise models and denoising algorithms, when there is no noise in the image or the image region is processed by the denoising algorithm, the original image information is often distorted, the image frequency domain information is filtered, and the image texture or feature points are lost, etc.

[0005] (2) The current image denoising using deep learning (such as PRIDNet, Noise2Noise, Nise2Void, etc.) is also a global image denoising process, and it is for common image noise, and the processing effect for nuclear irradiation noise is poor.

[0006] (3) The current image denoising method based on deep learning is generally divided into supervised learning and unsupervised learning. Using supervised learning to train the model requires noise and noise-free paired images in the same scene, but in reality it is difficult to obtain noise-free images in the same scene. Using unsupervised learning to train the model eliminates the need for noise-free data, but it also has the problem of large amount of training data and reduced denoising effect. SUMMARY

[0007] In order to solve the technical problems mentioned in the background art, a nuclear irradiation image denoising system based on supervised learning is proposed, which first uses an improved Unet algorithm to quickly and effectively identify and obtain nuclear radiation noise, and then uses Noise2Mask algorithm for noise removal.

[0008] In order to achieve the above object, the present application adopts the following technical scheme:

[0009] A nuclear irradiation image denoising system based on supervised learning includes a nuclear irradiation noise recognition module and an image denoising module, both of which are implemented through a deep learning network and trained through a supervised learning method to achieve the deep learning network for image denoising;

[0010] The deep learning network for nuclear irradiation noise recognition inputs a nuclear irradiation noise image and outputs a mask image corresponding to the nuclear irradiation noise image;

[0011] The deep learning network for image denoising inputs a nuclear irradiation noise image and a mask image, fills the noise area indicated by the mask image with color blocks, restores the image information from the noise peripheral area, and restores the image information of the image noise area marked by the mask image when restoring the image noise point information.

[0012] Further description of the above technical scheme:

[0013] The deep learning network for nuclear irradiation noise recognition is a Shallow-Unet network, and the Shallow-Unet network structure includes a down-sampling layer, a connection layer, and an up-sampling layer;

[0014] The down-sampling layer contains two sets of repeated structures, each set of repeated structure including two sets of 3x3 convolution layers, a nonlinear RELU layer, and a set of 2x2 max pooling layers;

[0015] The connection layer includes three sets of 3x3 convolution layers, a nonlinear RELU layer, and a set of CBAM spatial and channel mixed attention mechanisms;

[0016] The up-sampling layer contains two sets of repeated structures, the first set of repeated structures including three sets of 3x3 convolution layers, a nonlinear RELU layer, a set of CBAM spatial and channel mixed attention mechanisms, and a 2x2 deconvolution layer, and the second set of repeated structures including the same structure as the first set of repeated structures and a set of 1x1 convolution layers;

[0017] The corresponding two sets of repeated structures in the up-sampling and down-sampling layers are connected through a feature concatenation concat layer.

[0018] Further description of the above technical scheme:

[0019] The deep learning network for image denoising is a Noise2Mask network, which is obtained by training the unsupervised learning Noise2Void network using a supervised learning method.

[0020] As a further description of the above technical solution:

[0021] The Noise2Void network is trained by a supervised learning method to obtain the Noise2Mask network using the artificially generated noise image, the noise mask image and the noise-free image.

[0022] As a further description of the above technical solution:

[0023] The artificially generated noise image is a noise mask image corresponding to a noise picture output after randomly adding a pure color rectangular block with random quantity, random position and random size on a noise-free image dataset.

[0024] As a further description of the above technical solution:

[0025] The size of the pure color rectangular block is randomly selected within a range of [1, 20] pixels in width and height.

[0026] As a further description of the above technical solution:

[0027] When the Noise2Void network is trained by the supervised learning method to obtain the Noise2Mask network, the MSE is used as the loss function, and the function structure is:

[0028]

[0029] I is a noise-free image, K is a noise image, m and n are the width and height of the image, and i and j are the positions of the image.

[0030] As a further description of the above technical solution:

[0031] The mask image corresponding to the nuclear irradiation noise image marks the position and shape information of the input nuclear irradiation noise image.

[0032] As described above, due to the adoption of the above technical solution, the beneficial effects of the present application are:

[0033] 1、In the present application, the nuclear irradiation noise in the labeled image is learned by using a deep learning network, and the network model can automatically learn the complex, high-precision and robust noise features from the noise image, avoiding the work of artificial feature extraction.

[0034] 2、In the present application, only the area where the nuclear irradiation noise appears can be processed for noise reduction or image information recovery, and the image information in the noise-free area is not distorted.

[0035] 3、The training method of the nuclear radiation noise denoising model using the supervised learning method is improved, the problem of difficult acquisition of nuclear radiation noise and noise-free image pair data sets when using supervised learning is solved, the denoising problem of nuclear radiation noise with complex features is converted into the denoising problem of color blocks with simple features through feature recognition and accurate positioning of the nuclear radiation noise, so that a large number of noise and noise-free paired images can be generated manually for model training when training the denoising network model. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 An image processing flowchart of a nuclear radiation image denoising system based on supervised learning is shown according to an embodiment of the present application;

[0037] Figure 2 A Unet original network structure diagram provided by the present application is shown according to an embodiment of the present application;

[0038] Figure 3 A Shallow-Unet network structure diagram of a nuclear radiation image denoising system based on supervised learning is shown according to an embodiment of the present application;

[0039] Figure 4 A Shallow-Unet network model training flowchart of a nuclear radiation image denoising system based on supervised learning is shown according to an embodiment of the present application;

[0040] Figure 5 A denoising network training data set production flowchart of a nuclear radiation image denoising system based on supervised learning is shown according to an embodiment of the present application;

[0041] Figure 6 A Noise2Mask network model training flowchart of a nuclear radiation image denoising system based on supervised learning is shown according to an embodiment of the present application;

[0042] Figure 7 A nuclear radiation image noise diagram is shown according to an embodiment of the present application;

[0043] Figure 8 A comparison diagram of the effect before and after nuclear radiation noise image denoising is shown according to an embodiment of the present application. DETAILED DESCRIPTION

[0044] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0045] Embodiment one

[0046] Please refer to Figures 1-8 The present application provides a technical solution: a nuclear irradiation image denoising system based on supervised learning, which comprises a nuclear irradiation noise recognition module and an image denoising module, both of which are realized by a deep learning network, and the overall image processing flow is as shown in Figure 1 .

[0047] 1. Nuclear irradiation noise recognition module

[0048] (2) According to the size characteristics of nuclear irradiation noise, an improved UNet network Shallow-Unet is used to optimize the detection performance and efficiency of UNet for small size image targets;

[0049] (2) Through the training and learning of nuclear irradiation noise labeled data, the image features of nuclear irradiation noise are learned by Shallow-Unet network. After the model training is completed, the image with nuclear irradiation noise is input, and the Shallow-Unet network outputs the mask graph of nuclear irradiation noise. In the nuclear irradiation noise mask graph, the position and shape information of the nuclear irradiation noise in the input image are labeled.

[0050] Specifically, as shown in Figure 3 , the Shallow-Unet network structure includes down-sampling layer, connection layer and up-sampling layer;

[0051] The down-sampling layer contains two groups of repeated structures, each group of repeated structure includes two groups of 3x3 convolution layer (conv), nonlinear RELU layer and a group of 2x2 max pooling layer (stride = 2), and its main function is to extract features, and the number of feature channels is doubled after each down-sampling;

[0052] The connection layer includes three groups of 3x3 convolution layer, nonlinear RELU layer and a group of CBAM spatial and channel mixed attention mechanism;

[0053] The up-sampling layer includes two groups of repeated structures, each group of repeated structures including three groups of 3x3 convolution layers, a nonlinear RELU layer, and a group of CBAM spatial and channel mixed attention mechanisms, a 2x2 de-convolution layer, and the second group of repeated structures including the same structure as the first group of repeated structures and a group of 1x1 convolution layers, which mainly perform up-sampling. The up-sampling can make the low-resolution picture containing high-level abstract features become high-resolution while retaining the high-level abstract features. At each step, up-convolution is used first, and each time the use of up-convolution halves the number of feature channels and doubles the size of the feature map.

[0054] The corresponding two groups of repeated structures in the up-sampling and down-sampling layers are connected through a feature concatenation concat layer, and after de-convolution, the de-convolution result is concatenated with the feature map of the corresponding step in the down-sampling layer to perform feature fusion, so that the information becomes richer.

[0055] The feature map size in the down-sampling layer is slightly larger, and after pruning, it is spliced (here, the sizes of the two feature maps are adjusted to be consistent, and then spliced by channel number). The spliced map is then convolved twice with a 3x3 kernel, and the last layer has a 1x1 kernel size, converting the 64-channel feature map into a specific number of categories (classification number) of results.

[0056] 2. Image denoising module

[0057] (1) This module implements an image denoising Noise2Mask network based on the improved Noise2Void. The original Noise2Void network is an unsupervised learning network, and the model training only uses noise images. The improved Noise2Mask uses a supervised learning method for training, and uses artificially generated noise, noise mask images and noise-free images (ground truth) for model training. During training, the noise area indicated by the noise mask image in the noise image is pre-filled with color blocks.

[0058] (2) During inference, the input of the Noise2Mask model is: a) nuclear irradiation noise image b) nuclear irradiation noise image corresponding noise mask image. In the model image preprocessing, the nuclear irradiation noise area is filled with color blocks. With this method, the image denoising module simplifies the problem of denoising nuclear radiation noise to the problem of denoising simple color blocks, greatly facilitating the training of the denoising network.

[0059] (3) Noise2Mask network follows the basic assumption of Noise2Void: the pixels in the image are not independent, each pixel and the surrounding pixels are conditionally dependent, so the image information can be recovered from the noise surrounding area. However, unlike Noise2Void which performs global noise reduction on the image, Noise2Mask network uses the input of noise mask information to recover the image noise information, and only recovers the image information in the image noise area marked by the mask.

[0060] Based on the foregoing, the embodiment also provides a nuclear irradiation image denoising method based on supervised learning, and the implementation steps are as follows:

[0061] 1. Nuclear irradiation noise recognition network training:

[0062] (1) The original UNet network structure is as shown in Figure 2 The network structure is symmetrical, like the English letter U, so it is called UNet. UNet is an Encoder-Decoder structure, the first half is feature extraction, and the second half is up-sampling. The original Unet network has a 5-layer network structure to detect targets of different sizes in the image. The deeper the network, the greater the perception.

[0063] (2) In this embodiment, in order to improve the performance and efficiency of the nuclear irradiation noise detection module, the Unet network is modified to a 3-layer structure according to the characteristics of small nuclear irradiation noise size and few global features, and CBAM spatial and channel hybrid attention mechanism is added to Unet to enhance the network's feature extraction ability for nuclear radiation noise. The modified Shallow-Unet network structure is as shown in Figure 3 .

[0064] (3) Noise data labeling is performed on the nuclear irradiation noise dataset, and the training set, validation set and test set of the Shallow-Unet network are prepared.

[0065] (4) The labeled data set is used for model training, and the loss function is used to determine whether the training effect is reached. The test set is used to evaluate the learning effect of the model on the nuclear irradiation noise, and the model training process is as shown in Figure 4 .

[0066] 2. Image denoising network training:

[0067] (1) Training data set preparation. As shown in Figure 4As shown, on the noise-free data set (ground truth), the program randomly adds a number of random, location-random and width-height respectively located in [1, 20] pixel random size pure color rectangular color blocks. The image after adding the random color block is the noise picture of the training data set. The program also outputs the binary noise mask (mask) corresponding to the noise picture.

[0068] (2) The noise picture and the binary noise mask (mask) are used for network forward inference calculation, and the noise-free data set (ground truth) is used for evaluating the network prediction result.

[0069] (3) When training the Noise2Mask model, use MSE (mean square error calculation formula) as the loss function to evaluate the difference between the recovered image of the noise area of the model and the actual value (ground truth). The loss function structure is:

[0070]

[0071] I is a noise-free image, K is a noise image, m and n are the width and height of the image, and i and j are the positions of the image.

[0072] (4) The training process of the Noise2Mask model is as shown in Figure 5 .

[0073] In summary, in the embodiment, first, the nuclear radiation noise in the labeled image is learned using a deep learning network. The network model can automatically learn complex, high-precision and robust noise features from noise images, avoiding the work of manual feature extraction. In image processing applications, the use of noise features learned by the network can quickly and accurately identify whether the image contains nuclear radiation noise and the noise position;

[0074] Second, the noise reduction processing or image information recovery can be performed only on the area where the nuclear radiation noise appears, and the image information in the noise-free area is not distorted.

[0075] Finally, the embodiment improves the training method of the nuclear radiation noise denoising model using supervised learning, solving the problem of difficult access to data sets of nuclear radiation noise and noise-free images when using supervised learning. Through feature recognition and accurate positioning of nuclear radiation noise, the denoising problem of complex nuclear radiation noise is converted into a color block denoising problem with simple features. This makes it possible to quickly generate a large number of noise and noise-free paired images for model training during the training of the denoising network model. For example, Figure 8As shown, the effect comparison chart before and after the nuclear radiation noise image is denoised, wherein the left side is the nuclear radiation noise image, and the right side is the image processed by the denoising algorithm, and through the comparison of the images before and after the denoising, it can be seen that the application has a good nuclear radiation noise denoising effect.

[0076] The above is only a preferred embodiment of the application, but the protection scope of the application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the application within the technical range disclosed by the application, which should be covered in the protection scope of the application.

Claims

1. A supervised learning-based nuclear irradiation image denoising system, characterized in that, It includes a nuclear irradiation noise recognition module and an image denoising module. Both modules are implemented using deep learning networks, and the deep learning network for image denoising is trained using supervised learning methods. A deep learning network for identifying nuclear irradiation noise takes a nuclear irradiation noise image as input and outputs a mask image corresponding to the nuclear irradiation noise image. A deep learning network for image denoising is input with a nuclear irradiation noise image and a mask image. The noise regions in the noise image indicated by the mask image are filled with color blocks. Image information is recovered from the surrounding noise region. With the help of the mask image, image information is recovered from the image noise region marked by the mask image when recovering image noise information. The deep learning network used to identify nuclear irradiation noise is the Shallow-Unet network. The Shallow-Unet network structure includes a downsampling layer, a connection layer, and an upsampling layer. The downsampling layer contains two sets of repeating structures, each set of repeating structures including two sets of 3×3 convolutional layers, a nonlinear ReLU layer and a set of 2×2 max pooling layers; The connection layer includes three sets of 3×3 convolutional layers, a nonlinear ReLU layer, and a set of CBAM spatial and channel hybrid attention mechanisms; The upsampling layer contains two sets of repeating structures. The first set of repeating structures includes three sets of 3×3 convolutional layers, a nonlinear ReLU layer, a set of CBAM spatial and channel hybrid attention mechanisms, and a 2×2 deconvolutional layer. The second set of repeating structures includes the same structure as the first set of repeating structures, plus a set of 1×1 convolutional layers. The corresponding sets of repeating structures in the upsampling and downsampling layers are connected by a feature concat layer.

2. The nuclear irradiation image denoising system based on supervised learning according to claim 1, characterized in that, The deep learning network used to achieve image denoising is the Noise2Mask network, which is obtained by training the unsupervised Noise2Void network using a supervised learning method.

3. The nuclear irradiation image denoising system based on supervised learning according to claim 2, characterized in that, The Noise2Void network was trained using a supervised learning method with artificially generated noisy images, noise mask images, and noise-free images to obtain the Noise2Mask network.

4. The nuclear irradiation image denoising system based on supervised learning according to claim 3, characterized in that, The artificially generated noise image is generated by randomly adding a number of solid-color rectangular blocks of random quantity, position, and size to a noise-free image dataset, and then outputting a noise mask image corresponding to the noise image.

5. A supervised learning-based nuclear irradiation image denoising system according to claim 4, characterized in that, The size of a solid color rectangular block is a random range where the width and height are within [1, 20] pixels respectively.

6. A supervised learning-based nuclear irradiation image denoising system according to claim 2, characterized in that, When training the Noise2Void network using supervised learning to obtain the Noise2Mask network, MSE is used as the loss function, and the function structure is as follows: I represents the noise-free image, K represents the noisy image, m and n represent the image width and height, and i and j represent the image position.

7. A supervised learning-based nuclear irradiation image denoising system according to claim 1, characterized in that, The mask image corresponding to the nuclear irradiation noise image is marked with the position and shape information of the input nuclear irradiation noise image.

Images