Fixed pattern noise elimination method based on deep residual network
Through the fixed-mode noise cancellation method based on deep residual network, the problem of difficult to remove fixed-mode noise in images in strong radiation environments is solved, and efficient noise removal and image quality improvement is achieved, suitable for real-time processing on embedded AI platform .
Patent Information
- Application Number
- CN202411458654.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-18
- Publication Date
- 2025-02-18
AI Technical Summary
In a strong radiation environment, dynamic fixed mode noise often occurs in the image, and existing denoising methods are difficult to effectively remove, resulting in extremely poor image quality and affecting functional applications such as object detection.
The fixed-mode noise cancellation method based on deep residual network is adopted, and noise cancellation is performed through the trained fixed-mode noise cancellation network, and refined feature extraction is used to capture context information of different scales.
Effectively remove fixed mode noise generated in radiated environments, significantly improve image quality, improve the performance of functional applications such as object detection, and can be processed in real time on embedded AI development boards with limited resource constraints.
Smart Images
Figure CN119477739A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of image processing in a strongly irradiated environment, and in particular relates to a method, system, electronic device and computer-readable storage medium for eliminating fixed pattern noise based on a deep residual network. Background Art
[0002] The multi-physical field coupling factors in the red zone of the nuclear island are complex, especially in the strong radiation environment, it will cause β decay to emit β rays, and then γ decay to emit γ rays of 1.17MeV and 1.33MeV. γ-ray materials have strong penetrability and can easily enter the interior of the CMOS sensor to collide with atoms to produce energy exchange, triggering the total dose effect of ionizing radiation, and ultimately causing long-term dynamic fixed pattern noise in the image. However, it is difficult to completely remove this fixed pattern noise using traditional non-uniformity correction algorithms, and the extremely poor image quality seriously affects further functional applications such as target detection.
[0003] Since deep convolutional neural networks won the ImageNet competition in the field of image classification with an absolute advantage, the development of convolutional neural networks has been rapid. At present, convolutional neural networks have achieved results far exceeding traditional methods in the fields of image classification, super-resolution, and target detection. However, it is difficult to achieve significant results by directly applying the network structure in the classic image classification field to the field of image denoising, and the network structure is complex and the calculation amount of model parameters is very large.
[0004] In order to solve the above problems, the present invention proposes a fixed pattern noise removal method based on a deep residual network. By fusing the residual skip-layer connection structure, refined feature extraction is performed on each level of pooling layers to capture contextual information of different scales, thereby eliminating the fixed pattern noise in the image more efficiently and comprehensively. Summary of the invention
[0005] In order to solve the above-mentioned problems in the prior art, that is, to solve the problem that image degradation caused by radiation is common in camera equipment used in irradiated environments, and the existing denoising methods are difficult to remove the dynamic fixed pattern noise, and thus cannot obtain a clear image, the first aspect of the present invention proposes a fixed pattern noise removal method based on a deep residual network, the method comprising:
[0006] Obtain a degraded image collected in a real radiation environment as an input image;
[0007] Based on the input image, noise removal is performed by using a trained fixed pattern noise removal network to obtain a noise-free image;
[0008] Among them, the fixed pattern noise elimination network is constructed based on a deep learning residual network.
[0009] In some preferred embodiments, the fixed pattern noise elimination network includes a first convolution layer, a first pooling layer, N sequentially connected first CBR blocks, a second convolution layer, a second pooling layer, N sequentially connected second CBR blocks, a third convolution layer, a first deconvolution layer, a second deconvolution layer, a fourth convolution layer, a fifth convolution layer, a third deconvolution layer, and a sixth convolution layer; the first CBR block and the second CBR block both include sequentially connected convolution layers, batch normalization layers, and Relu activation function layers;
[0010] The first convolution layer, whose input is the input image; the first pooling layer, used for performing pooling processing on the output of the first convolution layer, and inputting N first CBR blocks connected in sequence after the processing;
[0011] The input of the second convolutional layer is the output of the last first CBR block among the N first CBR blocks connected in sequence; the second pooling layer is used to perform pooling processing on the output of the second convolutional layer, and input the processed output into the N second CBR blocks connected in sequence;
[0012] The input of the third convolutional layer is the output of the last second CBR block among the N second CBR blocks connected in sequence;
[0013] The input of the second deconvolution layer is the output of the third convolution layer; the output of the second deconvolution layer is concatenated with the output of the second convolution layer in the channel dimension and then input into the fourth convolution layer;
[0014] The input of the third deconvolution layer is the output of the fourth convolution layer;
[0015] The output of the second convolutional layer is processed by the first deconvolutional layer, and then concatenated with the output of the first convolutional layer in the channel dimension, and then input into the fifth convolutional layer after concatenation;
[0016] The output of the third deconvolution layer is concatenated with the output of the fifth convolution layer in the channel dimension and then input into the sixth convolution layer;
[0017] The input of the first convolutional layer is cross-multiplied with the output of the sixth convolutional layer, and the cross-multiplication result is used as the output of the fixed pattern noise removal network.
[0018] In some preferred embodiments, the fixed pattern noise elimination network is trained by:
[0019] Acquire an image set; perform data enhancement processing on the images in the image set, thereby expanding the image set;
[0020] The fixed pattern noise brought to the image by the radiation environment is simulated by using a pre-built fixed pattern noise polynomial model, and then the fixed simulated noise is combined with the image in the image set to obtain a noisy image;
[0021] Inputting the noisy image into a pre-constructed fixed pattern noise removal network to obtain a noise-removed image;
[0022] Calculating the loss value between the noise-removed image and its corresponding true value label, thereby updating the network parameters of the fixed pattern noise removal network;
[0023] The fixed pattern noise elimination network is trained cyclically until a trained fixed pattern noise elimination network is obtained.
[0024] In some preferred embodiments, the fixed pattern noise polynomial model is:
[0025]
[0026] Where S(i,j) represents fixed pattern noise, represents the polynomial coefficient of the jth column, Represents the image pixel value of the jth column.
[0027] In some preferred embodiments, the fixed pattern noise polynomial model is constructed by:
[0028] Collecting degraded images in a real radiation experimental environment; the real radiation experimental environment includes a radiation source, a dose rate at a camera location, and a cumulative total dose;
[0029] The fixed pattern noise characteristics of the acquired degraded images are analyzed in time and frequency, and a polynomial model is used to simulate the long-term dynamic fixed pattern noise.
[0030] The order of the polynomial is determined by combining the comprehensive algorithm complexity, algorithm complexity and fitting determination coefficient of different polynomials, and then the fixed pattern noise polynomial model is obtained.
[0031] In a second aspect of the present invention, a fixed pattern noise removal system based on a deep residual network is proposed, the system comprising:
[0032] An image acquisition module is configured to acquire a degraded image collected in a real radiation environment as an input image;
[0033] A noise removal module is configured to remove noise based on the input image by using a trained fixed pattern noise removal network to obtain a noise-free image;
[0034] Among them, the fixed pattern noise elimination network is constructed based on a deep learning residual network.
[0035] According to the third aspect of the present invention, an electronic device is proposed, comprising: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the above-mentioned fixed pattern noise elimination method based on a deep residual network.
[0036] In a fourth aspect of the present invention, a computer-readable storage medium is proposed, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by a computer to implement the above-mentioned fixed pattern noise elimination method based on a deep residual network.
[0037] Beneficial effects of the present invention:
[0038] The present invention can effectively remove the fixed pattern noise generated in the radiation environment and obtain a clear image.
[0039] 1) The present invention fully considers the characteristics of the fixed pattern noise in the radiation environment that changes dynamically with the accumulation of the total radiation dose, and fully removes the fixed pattern noise in the degraded image.
[0040] (2) The present invention can effectively extract noise features in degraded images, and effectively remove fixed pattern noise in radiation-degraded images through a fixed pattern noise elimination network constructed based on a deep residual network, thereby obtaining an ideal restored image, i.e., a clear image;
[0041] (3) Based on the good fixed pattern noise removal effect, the present invention uses model lightweight methods such as model compression to transplant and deploy it to an embedded AI development board with limited resource constraints on the end side, which can perform real-time processing. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings.
[0043] Figure 1 It is a flowchart of a method for eliminating fixed pattern noise based on a deep residual network according to an embodiment of the present invention;
[0044] Figure 2 It is a detailed flowchart of a method for eliminating fixed pattern noise based on a deep residual network according to an embodiment of the present invention;
[0045] Figure 3 is a schematic diagram of a deep residual network structure of an embodiment of the present invention;
[0046] Figure 4 This is a denoising result diagram of a grayscale image according to an embodiment of the present invention. DETAILED DESCRIPTION
[0047] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0048] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.
[0049] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application may be combined with each other.
[0050] A fixed pattern noise elimination method based on a deep residual network according to a first embodiment of the present invention is as follows: Figure 1 As shown, the following steps are included:
[0051] Obtain a degraded image collected in a real radiation environment as an input image;
[0052] Based on the input image, noise removal is performed by using a trained fixed pattern noise removal network to obtain a noise-free image;
[0053] Among them, the fixed pattern noise elimination network is constructed based on a deep learning residual network.
[0054] In order to more clearly illustrate the fixed pattern noise elimination method based on deep residual network of the present invention, the following is combined with the attached Figure 2 Each step in an embodiment of the method of the present invention is described in detail.
[0055] At present, the image degradation problem caused by radiation is common in the camera equipment used in the irradiated environment, and the existing denoising algorithm is difficult to solve this dynamic fixed pattern noise. Therefore, the present invention proposes a fixed pattern noise removal method based on a deep residual network, which can be processed in real time in the embedded AI platform on the end side to solve the above problem.
[0056] In the following embodiments, the training process of the fixed pattern noise elimination network is first described in detail, and then the process of noise elimination by the fixed pattern noise elimination method based on the deep residual network is described in detail.
[0057] 1. Training process of fixed pattern noise removal network
[0058] In this embodiment, a degraded image under a radiation environment is first collected, and its fixed pattern noise characteristics are analyzed in time and frequency, the generation mechanism of the long-term dynamic fixed pattern noise is analyzed, and a degraded image noise model (fixed pattern noise polynomial model) is established. The details are as follows:
[0059] Collect degraded images under radiation environment and perform time-frequency analysis on the fixed pattern noise characteristics; the preferred real radiation experimental environment of the present invention is 60 Co radiation source (it will undergo β decay to emit β rays, and then γ decay to emit γ rays of 1.17MeV and 1.33MeV. γ rays have strong penetrability and can easily enter the CMOS sensor to collide with atoms to produce energy exchange, triggering the total dose effect of ionizing radiation, and ultimately causing long-term dynamic fixed pattern noise in the image). The dose rate at the camera's location is 30Gy / h. When the camera is powered on, a degraded image is acquired in a continuous radiation environment, with a total accumulated dose of 2000Gy.
[0060] The generation mechanism of long-term dynamic fixed pattern noise is analyzed, and a degraded image noise model is established. A polynomial model is used to fit the fixed pattern noise. Assuming that V(i,j) represents the infrared output (i.e., image pixel value) of the (i,j)th detector on the CMOS image sensor, the fixed pattern noise S(i,j) can be expressed as:
[0061] S(i,j)=G(V(i,j)) (1)
[0062] Wherein, G represents the nonlinear mapping function of the fixed pattern noise.
[0063] According to the manifestation of fixed pattern noise, a polynomial is used to simulate the fixed pattern noise existing in long-term dynamics. The fixed pattern noise of the jth column can be expressed as:
[0064]
[0065] in: Represents the polynomial coefficient of the jth column. By adjusting the value of the polynomial coefficient, the polynomial model can simulate fixed pattern noise of different types and intensities. The determination coefficient R of the comprehensive algorithm complexity and algorithm complexity and fitting in this method is 2 ,Finally, a third-order polynomial is used to simulate the fixed pattern noise brought to the image by the radiation environment.
[0066] Acquire an image set; perform data enhancement processing on the images in the image set, thereby expanding the image set;
[0067] In this embodiment, the source of the image set is preferably 1,000 pictures of daily life scenes collected by PLUG617R, including cars, pedestrians, and laboratory scenes. Then, ten data enhancement methods such as flipping, rotation, scaling, brightness and contrast adjustment are used to expand the data set to 10,000 pictures. Then, a noise model is used to establish a noisy-noiseless data pair, which is divided into a training set and a validation set.
[0068] The fixed pattern noise brought to the image by the radiation environment is simulated by using a pre-built fixed pattern noise polynomial model, and then the fixed simulated noise is combined with the image in the image set to obtain a noisy image;
[0069] Inputting the noisy image into a pre-constructed fixed pattern noise removal network to obtain a noise-removed image;
[0070] In traditional deep neural networks, the output of each layer can be expressed as:
[0071] y=F(x,W) (3)
[0072] Where: x is the input, W is the weight parameter of the layer, and F(x,W) is the output obtained by combining the linear transformation (convolution) of the current layer and the nonlinear activation function.
[0073] Using ResNet residual connection, the output form of each layer is modified as follows:
[0074] y l =F(x,W)+x l (4)
[0075] Where: x l It is a skip connection in the residual connection, that is, directly inputting x l Added to the transformed output, the output y l It no longer depends solely on the convolution transformation of the current layer, but includes an identity mapping x l This design is more conducive to the extraction of noise features and avoids the problem of gradient disappearance.
[0076] Therefore, in this embodiment, the fixed pattern noise elimination network is based on the deep learning residual network, and the final noise extraction also uses the idea of residual jump connection, which fuses the input with the noise features extracted by the deep learning network in the channel dimension, and then generates the final noise feature map. Specifically: the fixed pattern noise elimination network includes a first convolutional layer, a first pooling layer, N first CBR blocks connected in sequence, a second convolutional layer, a second pooling layer, N second CBR blocks connected in sequence, a third convolutional layer, a first deconvolutional layer, a second deconvolutional layer, a fourth convolutional layer, a fifth convolutional layer, a third deconvolutional layer, and a sixth convolutional layer; the first CBR block and the second CBR block both include a convolutional layer, a batch normalization layer, and a Relu activation function layer (convolutional layer, i.e., a convolutional layer connected in sequence). Figure 3 The Conv in the pooling layer is Figure 3 The Pooling, Relu activation function layer in Figure 3 Relu in the deconvolution layer Figure 3 Deconv in the channel dimension is Concat);
[0077] The first convolution layer, whose input is the input image (i.e., the degraded image); the first pooling layer, used for performing pooling processing on the output of the first convolution layer, and inputting N first CBR blocks connected in sequence after the processing;
[0078] The input of the second convolutional layer is the output of the last first CBR block among the N first CBR blocks connected in sequence; the second pooling layer is used to perform pooling processing on the output of the second convolutional layer, and input the processed output into the N second CBR blocks connected in sequence;
[0079] The input of the third convolutional layer is the output of the last second CBR block among the N second CBR blocks connected in sequence;
[0080] The input of the second deconvolution layer is the output of the third convolution layer; the output of the second deconvolution layer is concatenated with the output of the second convolution layer in the channel dimension and then input into the fourth convolution layer;
[0081] The input of the third deconvolution layer is the output of the fourth convolution layer;
[0082] The output of the second convolutional layer is processed by the first deconvolutional layer, and then concatenated with the output of the first convolutional layer in the channel dimension, and then input into the fifth convolutional layer after concatenation;
[0083] The output of the third deconvolution layer is concatenated with the output of the fifth convolution layer in the channel dimension and then input into the sixth convolution layer;
[0084] The input of the first convolutional layer is cross-multiplied with the output of the sixth convolutional layer, and the cross-multiplication result is used as the output of the fixed pattern noise removal network.
[0085] Calculating the loss value between the noise-removed image and its corresponding true value label, thereby updating the network parameters of the fixed pattern noise removal network;
[0086] The fixed pattern noise elimination network is trained cyclically until a trained fixed pattern noise elimination network is obtained.
[0087] After training, the test set (the test set is a degraded image obtained in a real radiation environment) is used to test on a simulated noise image that is different from the common ones. The denoising performance indicators such as MSE and PNSR on this test set are extremely high. While removing noise, the image detail information is retained to the greatest extent.
[0088] To prove the effectiveness of this method, it is compared with traditional non-uniformity correction methods and a classic deep learning denoising method. Traditional non-uniformity correction methods include time-domain high-pass filtering and median histogram equalization (MIRE) algorithm. The classic deep learning denoising algorithm is the DLS-NUC method. In order to measure the performance of different methods, a group of degraded images with different fixed pattern noise levels under irradiation environment are selected, including 10 degraded images with target imaging and without targets, and the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) with the clear image are calculated respectively. SSIM and PSNR are indicators used to measure the degree of distortion of the denoised image. The larger the value, the smaller the image distortion. As shown in Table 1:
[0089] Table 1 Performance Indicators Time domain high pass filtering MIRE DLS-NUC Method of the present invention PSNR(dB) 10.33 13.28 24.23 36.67 SSIM(0-1) 0.58 0.73 0.79 0.89
[0090] From the above results, it can be seen that the method of the present invention is significantly better than other algorithms, especially the traditional non-uniformity correction method. Even compared with the advanced DLS-NUC algorithm, the method of the present invention still has a 12.4dB lead in PSNR and a 0.1 lead in SSIM. This proves that the method of the present invention has outstanding denoising performance in irradiation-degraded images. The denoising results are shown in Figure 2. Figure 4 shown.
[0091] 2. Fixed pattern noise removal method based on deep residual network
[0092] Obtain a degraded image collected in a real radiation environment as an input image;
[0093] Based on the input image, noise removal is performed by using a trained fixed pattern noise removal network to obtain a noise-free image;
[0094] In this embodiment, the actual application process of the trained fixed pattern noise removal network is: the noisy image (ie, the degraded image) is directly input into the trained fixed pattern noise removal network for denoising to obtain a noise-free image, ie, a corresponding clear image.
[0095] In addition, lightweight model methods such as model compression can be used to optimize the inference speed while preserving the denoising performance of the method as much as possible, so that the deep learning fixed pattern denoising method can not only be processed in real time on a high-performance, resource-rich PC, but also be transplanted and deployed on an embedded AI development board with limited resources. It can also perform denoising in real time, reaching 36 frames per second, which meets real-time applications.
[0096] A fixed pattern noise elimination system based on a deep residual network according to a second embodiment of the present invention comprises:
[0097] An image acquisition module is configured to acquire a degraded image collected in a real radiation environment as an input image;
[0098] A noise removal module is configured to remove noise based on the input image by using a trained fixed pattern noise removal network to obtain a noise-free image;
[0099] Among them, the fixed pattern noise elimination network is constructed based on a deep learning residual network.
[0100] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.
[0101] It should be noted that the fixed pattern noise elimination system based on the deep residual network provided in the above embodiment is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be decomposed or combined. For example, the modules in the above embodiments can be combined into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the modules or steps and are not regarded as improper limitations of the present invention.
[0102] An electronic device of the third embodiment of the present invention includes: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the above-mentioned fixed pattern noise elimination method based on a deep residual network.
[0103] A computer-readable storage medium according to a fourth embodiment of the present invention stores computer instructions, and the computer instructions are used to be executed by the computer to implement the above-mentioned fixed pattern noise elimination method based on a deep residual network.
[0104] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the electronic device and computer-readable storage medium described above can refer to the corresponding process in the aforementioned method example and will not be repeated here.
[0105] Those skilled in the art should be able to appreciate that the modules and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, computer software or a combination of the two, and the programs corresponding to the software modules and method steps can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the technical field. In order to clearly illustrate the interchangeability of electronic hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are performed in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.
[0106] The terms "first", "second", etc. are used to distinguish similar objects rather than to describe or indicate a particular order or sequence.
[0107] The term "comprise" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that includes a list of elements includes not only those elements but also other elements not expressly listed, or also includes elements inherent to such process, method, article, or apparatus / device.
[0108] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.
Claims
1. A fixed pattern noise elimination method based on deep residual network, characterized in that: The method comprises the following steps: Obtain a degraded image collected in a real radiation environment as an input image; Based on the input image, noise removal is performed by using a trained fixed pattern noise removal network to obtain a noise-free image; Among them, the fixed pattern noise elimination network is constructed based on a deep learning residual network.
2. According to the method for eliminating fixed pattern noise based on deep residual network in claim 1, it is characterized in that: The fixed pattern noise elimination network includes a first convolution layer, a first pooling layer, N first CBR blocks connected in sequence, a second convolution layer, a second pooling layer, N second CBR blocks connected in sequence, a third convolution layer, a first deconvolution layer, a second deconvolution layer, a fourth convolution layer, a fifth convolution layer, a third deconvolution layer, and a sixth convolution layer; the first CBR block and the second CBR block both include a convolution layer, a batch normalization layer, and a Relu activation function layer connected in sequence; The first convolution layer, whose input is the input image; the first pooling layer, used for performing pooling processing on the output of the first convolution layer, and inputting N first CBR blocks connected in sequence after the processing; The input of the second convolutional layer is the output of the last first CBR block among the N first CBR blocks connected in sequence; the second pooling layer is used to perform pooling processing on the output of the second convolutional layer, and input the processed output into the N second CBR blocks connected in sequence; The input of the third convolutional layer is the output of the last second CBR block among the N second CBR blocks connected in sequence; The input of the second deconvolution layer is the output of the third convolution layer; the output of the second deconvolution layer is concatenated with the output of the second convolution layer in the channel dimension and then input into the fourth convolution layer; The input of the third deconvolution layer is the output of the fourth convolution layer; The output of the second convolutional layer is processed by the first deconvolutional layer, and then concatenated with the output of the first convolutional layer in the channel dimension, and then input into the fifth convolutional layer after concatenation; The output of the third deconvolution layer is concatenated with the output of the fifth convolution layer in the channel dimension and then input into the sixth convolution layer; The input of the first convolutional layer is cross-multiplied with the output of the sixth convolutional layer, and the cross-multiplication result is used as the output of the fixed pattern noise removal network.
3. According to claim 2, a method for eliminating fixed pattern noise based on a deep residual network is characterized in that: The fixed pattern noise elimination network is trained by: Acquire an image set; perform data enhancement processing on the images in the image set, thereby expanding the image set; The fixed pattern noise brought to the image by the radiation environment is simulated by using a pre-built fixed pattern noise polynomial model, and then the fixed simulated noise is combined with the image in the image set to obtain a noisy image; Inputting the noisy image into a pre-constructed fixed pattern noise removal network to obtain a noise-removed image; Calculating the loss value between the noise-removed image and its corresponding true value label, thereby updating the network parameters of the fixed pattern noise removal network; The fixed pattern noise elimination network is trained cyclically until a trained fixed pattern noise elimination network is obtained.
4. According to claim 3, a method for eliminating fixed pattern noise based on deep residual network is characterized in that: The data enhancement processing includes flipping, rotation, scaling, brightness and contrast adjustment.
5. According to claim 3, a method for eliminating fixed pattern noise based on deep residual network is characterized in that: The fixed pattern noise polynomial model is: Where S(i,j) represents fixed pattern noise, represents the polynomial coefficient of the jth column, Represents the image pixel value of the jth column.
6. The method for eliminating fixed pattern noise based on deep residual network according to claim 5, characterized in that: The fixed pattern noise polynomial model is constructed by: Collecting degraded images in a real radiation experimental environment; the real radiation experimental environment includes a radiation source, a dose rate at a camera location, and a cumulative total dose; The fixed pattern noise characteristics of the acquired degraded images are analyzed in time and frequency, and a polynomial model is used to simulate the long-term dynamic fixed pattern noise. The order of the polynomial is determined by combining the comprehensive algorithm complexity, algorithm complexity and fitting determination coefficient of different polynomials, and then the fixed pattern noise polynomial model is obtained.
7. A fixed pattern noise removal system based on deep residual network, characterized in that: The system includes: An image acquisition module is configured to acquire a degraded image collected in a real radiation environment as an input image; A noise removal module is configured to remove noise based on the input image by using a trained fixed pattern noise removal network to obtain a noise-free image; Among them, the fixed pattern noise elimination network is constructed based on a deep learning residual network.
8. An electronic device, characterized in that: include: at least one processor; And a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement a fixed pattern noise elimination method based on a deep residual network as described in any one of claims 1-6.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by a computer to implement a fixed pattern noise elimination method based on a deep residual network as described in any one of claims 1-6.