A radiation interference image restoration method based on multi-scale attention mechanism

Abstract

The present invention proposes a method for restoring radiation interference images based on a multi-scale attention mechanism. A data set is created based on the characteristics of radiation interference degraded images to train an image restoration network model; the network model includes two parallel parts: a first feature extraction channel and a second feature extraction channel. The present invention uses a dual-channel network structure to simultaneously extract different levels of features in the image and introduces a multi-scale attention mechanism into the network model. That is, the noise information extracted by the CNN sub-network that detects noise features in the image in the second feature extraction channel is fused through the encoding structure of the U-Net network and the first feature extraction channel, fully tapping the extracted noise features and utilizing the different levels of features in severely degraded areas to stimulate the potential of the network, thereby maximizing the utilization of the extracted noise feature prior information and improving the image reconstruction quality.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a radiation interference image restoration method based on a multi-scale attention mechanism. Background Art

[0002] With the rapid development of computer vision technology, more and more industries are leveraging visual technology to solve related problems. However, due to the inherent physical limitations of image acquisition equipment, the image acquisition process is easily contaminated by various noises, especially in fields such as medicine, remote sensing imaging, and the nuclear industry, which are exposed to radioactive environments. Images in these scenarios are subject to interference from nuclear radiation, causing image degradation and, in severe cases, permanent damage to the electronic components within visual sensor devices such as cameras. Images affected by radiation interference severely affect image quality and effective information extraction, hindering the effectiveness of various visual technologies in radioactive environments. Therefore, radiation interference image restoration technology is crucial to improving the effectiveness of various visual technologies in radioactive environments. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention provides a radiation interference image restoration method based on a multi-scale attention mechanism, the purpose of which is to improve the restoration effect of images interfered with by radiation.

[0004] The present invention is achieved through the following technical solutions:

[0005] A radiation interference image restoration method based on a multi-scale attention mechanism includes the following steps:

[0006] (1) Obtaining a training data set: constructing a pseudo-radiated degraded image from the acquired image that is not disturbed by radiation according to the characteristics of the radiation disturbed image; forming an image pair from the image that is not disturbed by radiation and the degraded image that is pseudo-radiated; constructing a noise attention map from the noise position and noise intensity obtained from the degraded image that is pseudo-radiated; forming an image group from the original image that is not disturbed by radiation, the degraded image that is pseudo-radiated, and the noise attention map; then repeatedly constructing several image groups from several images that are not disturbed by radiation according to the above method, unifying the resolution of the images in the image group, and preprocessing each image in the image group into image blocks of equal size to obtain a training data set;

[0007] (2) Establishing an image restoration network model with a multi-scale attention mechanism: the image restoration network model is composed of a first feature extraction channel and a second feature extraction channel in parallel; the first feature extraction channel is composed of a U-Net network, and the second feature extraction channel is composed of a noise feature detection network and an encoding structure part of the U-Net network; the multi-scale attention mechanism is specifically to superimpose the output feature map of each module in the encoding structure part of the U-Net network of the first feature channel and the output feature map of the same scale module in the encoding structure part of the U-Net network of the second feature extraction channel point by point, as the input of the next module of the encoding structure part of the U-Net network of the first feature channel; the image restoration network model is trained using the training data set described in step (1) to obtain a trained image restoration model;

[0008] (3) Restoring the radiation interference image: The radiation interference image is input into the image restoration network model of the multi-scale attention mechanism established in step (2), and the output is a clear and noise-free image;

[0009] Furthermore, the noise attention map in step (1) is an image containing the position and intensity of noise points in the degraded image obtained by performing a morphological operation on the degraded image after pseudo-radiation according to the characteristics of the image noise after radiation interference; specifically, the image I obtained by performing an opening operation on the degraded image after pseudo-radiation and the degraded image after pseudo-radiation noiseopen Subtraction to obtain bright speckle noise position map I noiselight ; Image I obtained by inverting the degraded image after pseudo-radiation noiseinv and for image I noiseinv Image I after opening operation noiseinvopen Subtraction to obtain the noise position map of dark spots I noisedark , image I noiselight and image I noisedark After superposition, the noise attention map is obtained.

[0010] Furthermore, in the first feature extraction channel in step (2), the U-Net network adopts sub-pixel convolution operation for upsampling, the jump connection adopts the method of superposition of corresponding elements of tensors to realize information fusion, and the convolution layers with channel numbers of 512 and 1024 in the U-Net network are deleted, so that the number of downsampling and upsampling in the U-Net network is reduced to 2 times.

[0011] Furthermore, the second feature extraction channel in step (2) is composed of a noise feature detection network and an encoding structure part of a U-Net network, and the two modules are connected in sequence.

[0012] Furthermore, the noise feature detection network in the second feature extraction channel in step (2) is composed of five convolution modules, and on the basis of each convolution module being connected in sequence, the input of the first convolution module is directly connected to the fifth convolution module;

[0013] Specifically, the first convolution module consists of one convolution layer, the filter kernel size is 3×3×c, c represents the number of channels of the input image, when c=1, it means that the input image is a grayscale image, when c=3, it means that the input image is a color image; the second convolution module is composed of the residual block in the ResNet network; the third convolution module is composed of the dense connection module in the dense connection network; the fourth module is composed of the residual module in the residual network (ResNet) network; the fifth convolution module is composed of one convolution layer, the filter kernel size is 3×3×k, k represents the number of channels of the fourth module output feature map, and the number of channels of the output feature map of this convolution layer is c.

[0014] Furthermore, the encoding structure of the U-Net network in the second feature extraction channel in step (2) is as follows: the upsampling structure of the U-Net network and the convolutional layers with channel numbers of 512 and 1024 are deleted, so that the number of downsampling of the encoding structure of the U-Net network is reduced to 2 times.

[0015] Furthermore, the multi-scale attention mechanism in step (2) is specifically that the convolutional layer output feature map after each downsampling of the second feature extraction channel and the downsampling corresponding convolutional layer output feature map in the U-Net network of the first feature extraction channel are superimposed point by point as the input of the next convolutional layer of the U-Net network of the first feature extraction channel.

[0016] Furthermore, during the training process of the image restoration network model in step (2), the loss function used when training the model using the training data set is The expression is:

[0017]

[0018] in, Indicates the operation of finding the mean square error of two variables, represents the image reconstructed by the U-Net network decoding structure in the first feature extraction channel, I clear represents an image in the training dataset that is not affected by radiation interference, λ represents a hyperparameter that controls the proportion of the noise feature detection network loss in the second feature extraction channel to the overall loss, represents the output image of the noise feature detection network in the second feature extraction channel, I noiseattentionmap Represents the noise attention map in the training dataset image group.

[0019] The beneficial effects of the present invention are as follows:

[0020] A dual-channel network architecture is used to simultaneously extract information from different layers of the image. One channel constructs a CNN model to extract noise features. This transforms the task of acquiring noise location and intensity in a noisy image into detecting noise features within the image, which are then used as a noise attention map, simplifying the task. Specifically, a sub-network is built to learn how to remove image background and extract large noise specks. Because large noise specks are sparse in images, this approach significantly accelerates model convergence and quickly obtains image noise features. The U-Net network encoding structure fully exploits the noise information. A multi-scale attention mechanism is then used to fuse information with the U-Net network encoding structure of the other channel. This fully utilizes the different layers of image noise information to stimulate the network's potential and maximize the use of prior information on the extracted noise features. Finally, the decoding structure of the U-Net network of the other channel restores a clear image. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0022] Figure 1 Schematic diagram of the image restoration model training method flow of the present invention;

[0023] Figure 2 A schematic diagram of the noise attention map acquisition process of the present invention;

[0024] Figure 3 Schematic diagram of the image restoration model structure of the present invention;

[0025] Figure 4 Schematic diagram of the noise feature detection network (NoiseLocation-Net) of the present invention;

[0026] Figure 5 This is a functional principle block diagram of the image restoration method of the present invention. DETAILED DESCRIPTION

[0027] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention and are not to be construed as limiting the present invention.

[0028] Those skilled in the art will understand that unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art to which this invention belongs.

[0029] It should also be understood that terms, such as those defined in commonly used dictionaries, should be understood to have a meaning consistent with their meaning in the context of the prior art and will not be interpreted in an idealized or overly formal sense unless as defined herein.

[0030] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a," "an," "said," and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the stated features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, and / or groups thereof.

[0031] In the description of this specification, reference to the terms "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. In addition, those skilled in the art may combine and integrate different embodiments or examples described in this specification, as well as features of different embodiments or examples, unless otherwise inconsistent.

[0032] To facilitate understanding of the present invention, the present invention is further explained below with reference to specific embodiments in conjunction with the accompanying drawings. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0033] Those skilled in the art should understand that the drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily necessary for implementing the present invention.

[0034] Example 1

[0035] This embodiment 1 provides a system for training a model for image restoration subjected to radiation interference, the system comprising:

[0036] The acquisition module is used to obtain the training data set.

[0037] The training data set includes multiple image groups, each of which includes a degraded image after pseudo-radiation, an image not disturbed by radiation, and a noise attention map. The degraded image after pseudo-radiation is a degraded image constructed by pseudo-radiation interference, and the image not disturbed by radiation is an undegraded image under the same background. The noise attention map is a noise attention map that shows the noise position and noise intensity of the degraded image disturbed by pseudo-radiation, such as Figure 2 As shown, the degraded image after pseudo-radiation is subjected to an opening operation to obtain the image I noiseopen Specifically, the 3×3 sized structure element is first used to perform an erosion operation on the degraded image after pseudo radiation, and then the 3×3 sized structure element is used to perform an expansion operation on the degraded image after pseudo radiation; the degraded image after pseudo radiation and image I noiseopen Subtraction to obtain bright speckle noise position map I noiselight ; After negating the degraded image after pseudo-radiation, image I is obtained noiseinv , for image I noiseinv Perform an opening operation to obtain image I noiseinvopen Specifically, we first use a 3×3 size structure element to image I noiseinv Perform the corrosion operation and then use the 3×3 size structure element to image I noiseinv Do the dilation operation; noiseinv and image I noiseinvopen Subtraction to obtain the noise position map of dark spots I noisedark , image I noiselight and image I noisedark After superposition, the noise attention map is obtained;

[0038] Establish a module for building an image restoration network model.

[0039] like Figure 3 As shown, the established basic network model includes a first feature extraction channel and a second feature extraction channel in parallel, and the outputs of the first feature extraction channel and the second feature extraction channel are connected through a multi-scale attention mechanism;

[0040] The training module is used to train the image restoration network model using the training data set to obtain a trained image restoration model.

[0041] In this embodiment 1, the image restoration network model training system is used to implement the image restoration model training method, and finally the image restoration model is obtained. Figure 1 As shown, the method includes:

[0042] A training data set is acquired using an acquisition module. The training data set includes multiple image groups, each of which includes a degraded image after pseudo-radiation, an image not disturbed by radiation, and a noise attention map. The degraded image after pseudo-radiation is a degraded image disturbed by pseudo-radiation, the image not disturbed by radiation is an undegraded image under the same background, and the noise attention map is a noise attention map that displays the location and intensity of noise points in the degraded image after radiation.

[0043] Establishing an image restoration network model using a building module. The established image restoration network model includes a first feature extraction channel and a second feature extraction channel connected in parallel, wherein outputs of the first feature extraction channel and the second feature extraction channel are connected via a multi-scale attention mechanism;

[0044] The image restoration network model is trained using the training data set using a training module to obtain a trained image restoration network model.

[0045] like Figure 3 As shown, in this embodiment 1, the U-Net network in the first feature extraction channel adopts sub-pixel convolution operation for upsampling, and the jump connection adopts the method of superposition of corresponding elements of tensors to realize information fusion, and also includes deleting the convolution layers with channel numbers of 512 and 1024 in the U-Net network, so that the number of downsampling and upsampling in the U-Net network is reduced to 2 times.

[0046] In this embodiment 1, the second feature extraction channel is composed of a noise feature detection network and an encoding structure part of a U-Net network, and the two modules are connected in sequence.

[0047] The noise feature detection network consists of five convolutional modules. After each convolutional module is connected sequentially, the input of the first convolutional module is directly connected to the fifth convolutional module. The first convolutional module consists of one convolutional layer with a filter kernel size of 3×3×c, where c represents the number of channels of the input image. When c=1, it represents a grayscale image, and when c=3, it represents a color image. The second convolutional module consists of a residual block from the ResNet network. The third convolutional module consists of a densely connected module from the DenseNet network. The fourth module consists of a residual module from the ResNet network. The fifth convolutional module consists of one convolutional layer with a filter kernel size of 3×3×k, where k represents the number of channels of the fourth module's output feature map. The number of channels of this convolutional layer's output feature map is c.

[0048] Among them, the encoding structure part of the U-Net network in the second feature extraction channel is to delete the upsampling part structure in the U-Net network, and also includes deleting the convolution layers with channel numbers of 512 and 1024, so that the number of downsampling times of the encoding structure part of the U-Net network is reduced to 2 times.

[0049] In this embodiment 1, the image restoration network model is trained using the training data set to obtain a trained image restoration model, including:

[0050] Inputting the pseudo-radiated degraded images in the plurality of image groups into the image restoration network model to obtain a plurality of output images of the U-Net network in the first feature extraction channel and a plurality of output images of the noise feature detection network in the second feature extraction channel;

[0051] constructing a content loss function based on a plurality of output images of the U-Net network in the first feature extraction channel and corresponding images not disturbed by radiation in the plurality of image groups;

[0052] constructing a noise attention map loss function based on the multiple output images of the noise feature detection network in the second feature extraction channel and the corresponding noise attention maps in the multiple image groups;

[0053] The content loss function and the noise attention map loss function are superimposed in a certain proportion to construct the complete image restoration network model loss function;

[0054] Optimizing the minimum value of the loss function of the image restoration network model to obtain optimized network parameters;

[0055] The image restoration network model is iteratively updated using the optimized network parameters to obtain a final trained image restoration model.

[0056] Example 2

[0057] like Figure 5 As shown, in this embodiment 2, an image denoising method is provided, the method comprising:

[0058] An acquisition module, which acquires a contaminated and degraded image in a radiation scene to be processed;

[0059] The processing module is used to process the acquired image to be processed using the restoration network obtained by the image restoration network model training method to obtain a restored image.

[0060] In this embodiment 2, a method for restoring an image interfered with by radiation is implemented using the above-mentioned image restoration system. The method includes: using an acquisition module to acquire an image to be processed, and using a processing module to process the image interfered with by radiation using a trained image restoration model to obtain a restored clear image.

[0061] In this embodiment 2, the trained image restoration network model is obtained using an image restoration model training system, which includes:

[0062] An acquisition module is configured to acquire a training dataset. The training dataset includes multiple image groups, each of which includes a degraded image after pseudo-radiation, an image not disturbed by radiation, and a noise attention map. The degraded image after pseudo-radiation is a constructed degraded image with pseudo-radiation interference, the image not disturbed by radiation is an undegraded image under the same background, and the noise attention map is a noise attention map that displays the location and intensity of noise points in the degraded image with pseudo-radiation interference.

[0063] Establishing a module for establishing an image restoration network model. The established image restoration network model includes a first feature extraction channel and a second feature extraction channel connected in parallel, wherein the outputs of the first feature extraction channel and the second feature extraction channel are connected via a multi-scale attention mechanism;

[0064] The training module is used to train the image restoration network model using the training data set to obtain a trained image restoration model.

[0065] In this embodiment 2, the image restoration model is obtained by using the above-mentioned image restoration network model training system, including:

[0066] A training data set is acquired using an acquisition module. The training data set includes multiple image groups, each of which includes a degraded image after pseudo-radiation, an image not disturbed by radiation, and a noise attention map. The degraded image after pseudo-radiation is a constructed degraded image after pseudo-radiation, the image not disturbed by radiation is an undegraded image under the same background, and the noise attention map is a noise attention map that displays the location and intensity of noise points in the degraded image disturbed by pseudo-radiation.

[0067] Establishing an image restoration network model using a building module. The established image restoration network model includes a first feature extraction channel and a second feature extraction channel connected in parallel, wherein outputs of the first feature extraction channel and the second feature extraction channel are connected via a multi-scale attention mechanism;

[0068] The image restoration network model is trained using the training data set using a training module to obtain a trained image restoration network model.

[0069] In this embodiment 2, the U-Net network in the first feature extraction channel uses sub-pixel convolution operation for upsampling, and the jump connection uses the method of superposition of corresponding elements of tensors to achieve information fusion. It also includes deleting the convolution layers with channel numbers of 512 and 1024 in the U-Net network, so that the number of downsampling and upsampling in the U-Net network is reduced to 2 times.

[0070] In this embodiment 2, the second feature extraction channel is composed of a noise feature detection network (NoiseLocation-Net) and an encoding structure part of a U-Net network, and the two modules are connected in sequence.

[0071] The noise feature detection network consists of five convolutional modules. After each convolutional module is connected sequentially, the input of the first convolutional module is directly connected to the fifth convolutional module. The first convolutional module consists of one convolutional layer with a filter kernel size of 3×3×c, where c represents the number of channels of the input image. When c=1, it represents a grayscale image, and when c=3, it represents a color image. The second convolutional module consists of a residual block from the ResNet network. The third convolutional module consists of a densely connected module from the DenseNet network. The fourth module consists of a residual module from the ResNet network. The fifth convolutional module consists of one convolutional layer with a filter kernel size of 3×3×k, where k represents the number of channels of the fourth module's output feature map. The number of channels of this convolutional layer's output feature map is c.

[0072] Among them, the encoding structure part of the U-Net network in the second feature extraction channel is to delete the upsampling part structure in the U-Net network, and also includes deleting the convolution layers with channel numbers of 512 and 1024, so that the number of downsampling times of the encoding structure part of the U-Net network is reduced to 2 times.

[0073] In this embodiment 2, the image restoration network model is trained using the training data set to obtain a trained image restoration model, including:

[0074] Inputting the pseudo-radiated degraded images in the plurality of image groups into the image restoration network model to obtain a plurality of output images of the U-Net network in the first feature extraction channel and a plurality of output images of the noise feature detection network in the second feature extraction channel;

[0075] constructing a content loss function based on a plurality of output images of the U-Net network in the first feature extraction channel and corresponding images not disturbed by radiation in the plurality of image groups;

[0076] constructing a noise attention map loss function based on the multiple output images of the noise feature detection network in the second feature extraction channel and the corresponding noise attention maps in the multiple image groups;

[0077] The content loss function and the noise attention map loss function are superimposed in a certain proportion to construct the complete image restoration network model loss function;

[0078] Optimizing the minimum value of the loss function of the image restoration network model to obtain optimized network parameters;

[0079] The image restoration network model is iteratively updated using the optimized network parameters to obtain a final trained image restoration model.

[0080] In this embodiment 2, the degradation equation I is used noise =HI clean +n to train the constructed image restoration network. Where H is the degradation matrix, HI clean is the low-resolution image obtained by the degradation matrix H after the clear image is obtained, n is the large noise introduced by the radiation single particle effect, I clean is the image without radiation interference, I noise is the radiation interference degraded image. The network model is used to predict I clean =f(I noise ) in the algorithm, and finally the real image degraded by radiation interference is input into the network model to obtain the restored clear image.

[0081] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solutions disclosed in the present invention without the need for creative work should be included in the scope of protection of the present invention.

Claims

1. A radiation interference image restoration method based on a multi-scale attention mechanism, characterized in that: The method comprises the following steps: (1) Obtaining a training data set: constructing a pseudo-radiated degraded image based on the characteristics of the radiation-interferenced image from the acquired image that is not interfered with by radiation; An image pair is formed by combining an image not disturbed by radiation and a degraded image after pseudo-radiation; The noise position and noise intensity of the degraded image after pseudo-radiance processing are used to construct a noise attention map; An image group is composed of the original image without radiation interference, the degraded image after pseudo-radiation, and the noise attention map; Then, several images that have not been disturbed by radiation are repeatedly constructed according to the above method to obtain several image groups. The resolution of the images in the image groups is unified, and each image in the image group is preprocessed into image blocks of equal size to obtain a training data set. (2) Establishing an image restoration network model with a multi-scale attention mechanism: the image restoration network model is composed of a first feature extraction channel and a second feature extraction channel in parallel; the first feature extraction channel is composed of a U-Net network, and the second feature extraction channel is composed of a noise feature detection network and an encoding structure part of the U-Net network; the multi-scale attention mechanism is specifically to superimpose the output feature map of each module in the encoding structure part of the U-Net network of the first feature channel and the output feature map of the same scale module in the encoding structure part of the U-Net network of the second feature extraction channel point by point, as the input of the next step of the encoding structure part of the U-Net network of the first feature channel; Using the training data set described in step (1) to train the image restoration network model to obtain a trained image restoration model; (3) Restoring the radiation interference image: The radiation interference image is input into the image restoration network model of the multi-scale attention mechanism established in step (2), and the output is a clear and noise-free image.

2. The radiation interference image restoration method based on the multi-scale attention mechanism according to claim 1 is characterized in that: The noise attention map in step (1) is an image containing the position and intensity of noise points in the degraded image obtained by performing a morphological operation on the degraded image after pseudo-radiation according to the characteristics of the image noise after radiation interference; specifically, the image I obtained by performing an opening operation on the degraded image after pseudo-radiation and the degraded image after pseudo-radiation. noiseopen Subtraction to obtain bright speckle noise position map I noiselight ; Image I obtained by inverting the degraded image after pseudo-radiation noiseinv and for image I noiseinv Image I after opening operation noiseinvopen Subtraction to obtain the noise position map of dark spots I noisedark , image I noiselight and image I noisedark After superposition, the noise attention map is obtained.

3. The radiation interference image restoration method based on the multi-scale attention mechanism according to claim 1 is characterized in that: In the first feature extraction channel in step (2), the U-Net network adopts sub-pixel convolution operation for upsampling, and the jump connection adopts the method of superposition of corresponding elements of tensors to realize information fusion, and deletes the convolution layers with channel numbers of 512 and 1024 in the U-Net network, so that the number of downsampling and upsampling in the U-Net network is reduced to 2 times.

4. The radiation interference image restoration method based on the multi-scale attention mechanism according to claim 1 is characterized in that: The second feature extraction channel in step (2) is composed of a noise feature detection network and an encoding structure part of a U-Net network, and the two modules are connected in sequence.

5. The radiation interference image restoration method based on the multi-scale attention mechanism according to claim 1 is characterized in that: The noise feature detection network in the second feature extraction channel in step (2) is composed of 5 convolution modules. On the basis of each convolution module being connected in sequence, the input of the first convolution module is directly connected to the fifth convolution module.

6. The radiation interference image restoration method based on the multi-scale attention mechanism according to claim 5 is characterized in that: The first convolution module consists of one convolution layer with a filter kernel size of 3×3×c, where c represents the number of channels of the input image. When c=1, it means that the input image is a grayscale image, and when c=3, it means that the input image is a color image. The second convolution module consists of the residual block in the ResNet network. The third convolution module consists of the dense connection module in the dense connection network. The fourth module consists of the residual module in the ResNet network. The fifth convolution module consists of one convolution layer with a filter kernel size of 3×3×k, where k represents the number of channels of the output feature map of the fourth module, and the number of channels of the output feature map of this convolution layer is c.

7. The radiation interference image restoration method based on the multi-scale attention mechanism according to claim 1 is characterized in that: The encoding structure part of the U-Net network in the second feature extraction channel in the step (2) is as follows: deleting the upsampling part structure and the convolution layers with channel numbers of 512 and 1024 in the U-Net network, so that the number of downsampling of the encoding structure part of the U-Net network is reduced to 2 times.

8. The radiation interference image restoration method based on the multi-scale attention mechanism according to claim 1 is characterized in that: The multi-scale attention mechanism in step (2) is specifically that the convolutional layer output feature map after each downsampling of the second feature extraction channel and the downsampling corresponding convolutional layer output feature map in the U-Net network of the first feature extraction channel are superimposed point by point as the input of the next convolutional layer of the U-Net network of the first feature extraction channel.

9. The radiation interference image restoration method based on the multi-scale attention mechanism according to claim 1 is characterized in that: During the training process of the image restoration network model in step (2), the loss function used when training the model using the training data set is The expression is: in, Indicates the operation of finding the mean square error of two variables, represents the image reconstructed by the U-Net network decoding structure in the first feature extraction channel, I clear represents an image in the training dataset that is not affected by radiation interference, λ represents a hyperparameter that controls the proportion of the noise feature detection network loss in the second feature extraction channel to the overall loss, represents the output image of the noise feature detection network in the second feature extraction channel, I noiseattentionmap Represents the noise attention map in the training dataset image group.

Images