Image denoising method for interference image

Abstract

The invention provides an image denoising method for an interference image. The method comprises the following steps: firstly, based on an interference principle, aiming at the problem that paired interference images are difficult to acquire, providing a process of randomly generating interference image pairs, generating interference simulation images by using a Legendre polynomial, adding random noise to simulate complex noise, and artificially increasing darkness to simulate real interference images; secondly, for the problem that the contrast ratio of the interference image is insufficient, the features of different channels of the interference image are obtained through triple feature extraction, and the contrast ratio of the image is remarkably improved; then, aiming at the problem that the interference image has multiple noises, a three-branch denoising network is used, the structure is the same but the input is different, the parameter quantity is reduced, meanwhile, denoising is carried out on different interference image features, and the texture details and the structure information of the image are better reserved; and finally, for detail loss possibly caused in the network, fusion modules are used at different joints, and the deployment of two times of fusion attention achieves a better feature integration effect.

Description

Technical Field:

[0001] The present invention relates to image processing technology, and specifically to an image denoising method for interference images. Technical Background:

[0002] Interference image denoising based on deep learning is an important research direction in the cross - field of computer vision and optics, aiming to improve the quality of interference images collected in practical applications. For interference images, traditional image denoising methods mainly rely on filtering techniques or statistical models, but these methods often perform poorly when dealing with complex noise or images with high noise levels. With the introduction of deep learning technology, the performance of image denoising has been significantly improved.

[0003] However, at present, there are few image denoising algorithms for interference images, which are still in the initial exploration stage. Interference images mainly have problems in terms of brightness and noise. Specifically, the following problems exist in the image denoising method for interference images: First, it is difficult to obtain interference images. There are no interference images in the existing public datasets, but most of the current enhancement algorithms are supervised methods, resulting in fewer algorithms specifically for interference images. Therefore, how to create an interference image dataset is one of the challenges we currently face. Second, the contrast of interference images is low. Existing low - illumination enhancement methods can improve the image brightness, but there are still over - exposure or under - exposure phenomena, and even saturation imbalance. This requires a multi - aspect enhancement processing method to enhance interference images from multiple angles. Third, interference images have a lot of noise. The large amount of transmission noise and the noise caused by the relatively dark interference images have an impact on the subsequent applications of interference images. Therefore, an effective denoising process is needed.

[0004] In response to the above problems of interference images, recently, RetinexFormer designed an end - to - end model with a single - stage framework, which can well enhance the image brightness. Qi et al. proposed the CloFormer network by using context - aware global and local enhancement. Global prior knowledge can capture the overall structural information of the image, while local prior knowledge can more finely process the details of the image. Wang et al. were inspired by dynamic filters and cascaded dynamic filters to propose CasDyF - Net. Although the existing technologies have made substantial progress, there is currently no available interference image dataset, and the problem of simultaneously processing the contrast and noise of interference images has not been solved. Therefore, designing an efficient interference image denoising method obviously has important research significance. Summary of the Invention:

[0005] The object of the present invention is to solve the problems of difficult acquisition of interference image datasets and insufficient contrast and excessive noise in interference images, and to propose an image denoising method for interference images. A dataset is made according to the interference formula and a relatively clean interference image is restored through a deep learning network.

[0006] To achieve the above object, the present invention provides an image denoising method for interference images, which mainly includes the following five parts: The first part is to make an interference image dataset and preprocess it; the second part is to perform multi-channel image enhancement and feature extraction on the interference image; the third part is to denoise the three image features obtained in the second part; the fourth part is to fuse the three features after denoising; the fifth part is the training and testing of the image denoising network model for interference images, and finally a relatively clean interference image is obtained. The specific steps are as follows:

[0007] The first part includes two steps:

[0008] Step 1: Generate interference simulation images by randomly selecting the order of the Legendre polynomial and randomly adjusting the phase coefficient, where the order is limited between 15 and 400. The interference simulation images are used as normal light images, and the brightness is artificially adjusted to obtain low-light images. A variety of types of local block noise and overall noise are randomly introduced into the dataset to obtain interference noise images. A total of 1000 groups of interference image pairs are generated to form an interference noise dataset, and it is divided into training data and test data according to a ratio of 4:1;

[0009] Step 2: Crop the interference image pairs into image blocks of the same size according to a sampling size of 128×128, and perform data augmentation operations on the sampled data to expand the data samples to form the final training set samples;

[0010] The second part includes thirteen steps;

[0011] Step 3: Input the training sample images in Step 2 into the network, and transform the feature L1 from the RGB space to the HSV space through channel transformation, and separate the three channels, namely the hue channel feature H1, the saturation channel feature S1, and the value channel feature V1;

[0012] Step 4: Use channel fusion on the hue channel feature H1 separated in Step 3 and the HSV feature L1 to obtain L2;

[0013] Step 5: After passing L2 in Step 4 through a 3×3 convolution and a 5×5 convolution, obtain the hue channel feature H2;

[0014] Step 6: Pass the hue channel feature H2 in Step 5 through a 3×3 convolution to obtain the hue channel feature H3;

[0015] Step 7: Obtain the enhanced saturation channel feature S2 from the saturation channel feature S1 separated in Step 3 through an adaptive block saturation enhancement function;

[0016] Step 8: Use channel fusion for the enhanced saturation channel feature S2 in Step 7 and the HSV feature L1 in Step 3 to obtain L3

[0017] Step 9: Obtain the saturation channel feature S3 after passing L3 in Step 8 through a 3×3 convolution and a 5×5 convolution;

[0018] Step 10: Obtain the saturation channel feature S4 by passing the saturation channel feature S3 in Step 9 through a 3×3 convolution;

[0019] Step 11: Use channel fusion for the value channel feature V1 separated in Step 3 and the HSV feature L1 to obtain L4;

[0020] Step 12: Obtain the value channel feature V2 after passing L4 in Step 11 through a 3×3 convolution and a 5×5 convolution;

[0021] Step 13: Obtain the value channel feature V3 by passing the value channel feature V2 in Step 12 through a 3×3 convolution;

[0022] Step 14: Based on a variant of the Retinex theory, perform a point-by-point multiplication operation on the value channel feature V3 obtained in Step 13 and the HSV feature L1 in Step 3 to obtain the value channel feature V4;

[0023] Step 15: Perform a pixel addition operation on the value channel feature V4 obtained in Step 14 and the HSV feature L1 in Step 3 to obtain the value channel feature V5;

[0024] The third part includes ten steps:

[0025] Step 16: Take the value channel feature V5 obtained in Step 15 and the saturation channel feature S4 obtained in Step 10 as inputs and put them into the fusion attention to obtain the enhanced image F1;

[0026] Step 17: Use convolution to perform channel expansion on the enhanced image F1 obtained in Step 16 and the hue channel feature H3 obtained in Step 6 to obtain the enhanced image F2 and the hue channel feature H4;

[0027] Step 18: Use the hue channel feature H4 in Step 17 as the input and the hue channel feature H2 in Step 5 as the guidance. After passing through the receptive field module RFB, the enhanced image G1.1 is obtained. The receptive field module RFB consists of a dual-branch attention, a residual depthwise separable convolutional block RDSCB, and a normalization operation LN. The dual-branch attention is formed by connecting the guidance attention DA and the global attention GA in parallel;

[0028] Step 19: Use the output feature G1.1 of Step 18 as the input of the receptive field module RFB to obtain the output feature G1.2;

[0029] Step 20: Use the output feature G1.2 of Step 19 as the input of the receptive field module RFB to obtain the output feature G1.3;

[0030] Step 21: Perform channel fusion on the output feature G1.2 of Step 19 and the output feature G1.3 of Step 20, and use it as the input of the receptive field module RFB to obtain the output feature G1.4;

[0031] Step 22: Perform channel fusion on the output feature G1.4 of Step 21 and the output feature G1.1 of Step 18, and use it as the input of the receptive field module RFB to obtain the output feature G1.5;

[0032] Step 23: Perform channel fusion on the output feature G1.5 of Step 22 and the hue channel feature H4 in Step 17 to obtain the output feature G1.6;

[0033] Step 24: Use the enhanced image F2 in Step 17 as the input and the saturation channel feature S3 in Step 9 as the guidance, and repeat the operations in Step 18 - Step 23 to obtain the enhanced image G2.n, where n = 1, 2, 3, 4, 5, 6;

[0034] Step 25: Use the enhanced image F2 in Step 17 as the input and the lightness channel feature V2 in Step 12 as the guidance, and repeat the operations in Step 18 - Step 23 to obtain the enhanced image G3.n, where n = 1, 2, 3, 4, 5, 6;

[0035] The fourth part includes one step:

[0036] Step 26: Use G1.6 obtained in Step 23, G2.6 obtained in Step 24, and G3.6 obtained in Step 25 as the input into the fusion attention to obtain the output image;

[0037] The fifth part includes two steps:

[0038] Step 27: Input the training set samples in Step 2 into the network from Step 3 to Step 26, and set the network hyperparameters: the learning rate is 2e-4, the batch size is 4, the optimizer is Adam, the loss function is L1Loss, and train the network to obtain the final pre-trained model for interferometric image denoising;

[0039] Step 28: Input the interferometric noise image test dataset obtained in Step 1 into the pre-trained model obtained in Step 27, and the network can recover the interferometric image without noise or with little noise.

[0040] The present invention provides an image denoising method for interferometric images. First, based on the interference principle, aiming at the difficulty in obtaining paired interferometric images, a method for randomly generating interferometric image pairs is proposed. The Legendre polynomial is used to generate interferometric simulation images, random noise is added to them to simulate complex noise, and their darkness is artificially increased to simulate real interferometric images. Secondly, aiming at the problem of insufficient contrast in interferometric images, the Retinex theory is adopted to enhance the brightness component of the interferometric image, and the adaptive saturation module is used to enhance the saturation component of the interferometric image. Through triple feature extraction, the features of different channels of the interferometric image are obtained, significantly improving the contrast of the image. Then, aiming at the problem of multiple noises in interferometric images, a three-branch denoising module is used, with the same structure but different inputs, reducing the number of parameters, denoising different interferometric image features, and better retaining the texture details and structural information of the image. Its denoising module uses multiple receptive field modules RFB to form an encoder-decoder network, and the dual-branch attention is composed of the guiding attention DA and the global attention GA connected in parallel, extracting feature representations under different receptive fields to capture multi-scale feature information. Subsequently, the shallow features are sent to the RDSCB module for deep feature enhancement. Finally, aiming at the possible loss of details in the network, a fusion module is used at different connections, and the deployment of the fusion attention twice achieves a better feature integration effect. Description of the Drawings:

[0041] Figure 1 It is the flowchart of dataset production of the present invention

[0042] Figure 2 It is the overall framework diagram of the network of the present invention;

[0043] Figure 3 It is the flowchart of adaptive saturation enhancement of the present invention;

[0044] Figure 4 It is the first fusion module Fusion of the present invention;

[0045] Figure 5 It is the receptive field module RFB of the present invention;

[0046] Figure 6 The second fusion module Fusion of the present invention;

[0047] Figure 7 is the interference noise image;

[0048] Figure 8 is the image processed by using the present invention Figure 7 afterwards. Specific implementation manner:

[0049] To better understand the present invention, the image denoising method of the present invention for interference images will be described in more detail below in conjunction with specific implementation manners. In the following descriptions, the detailed descriptions of the current prior art may dilute the subject matter of the present invention, and these descriptions will be ignored here.

[0050] The first part includes two steps:

[0051] Step 1: Generate an interference simulation image by randomly selecting the order of the Legendre polynomial and randomly adjusting the phase coefficient, where the order is restricted between 15 and 400. Take the interference simulation image as the normal light image, and introduce various types of local block noises and overall noises into the dataset randomly to obtain the interference noise image, and artificially adjust the brightness to obtain the low-light image. As Figure 1 shown, a total of 1000 groups of interference image pairs are generated, made into an interference noise dataset, and taken as training data and test data according to a ratio of 4:1;

[0052] Step 2: Crop the interference image pairs into image blocks of the same size according to a sampling size of 128×128, and perform data augmentation operations on the sampled data to expand the data samples to form the final training set samples;

[0053] Figure 2 is the overall network framework diagram of the image denoising method of the present invention for interference images. In this implementation, it is carried out according to the following steps:

[0054] The second part includes thirteen steps;

[0055] Step 3: Input the training sample images in Step 2 into the network, and transform the feature L1 from the RGB space to the HSV space through a channel transformation and separate the three channels, namely the hue channel feature H1, the saturation channel feature S1, and the value channel feature V1;

[0056] Step 4: Use channel fusion for the hue channel feature H1 separated in Step 3 and the HSV feature L1 to obtain L2;

[0057] Step 5: Pass L2 in Step 4 through a 3×3 convolution and a 5×5 convolution to obtain the hue channel feature H2;

[0058] Step 6: Convolve the hue channel feature H2 in Step 5 with a 3×3 convolution kernel to obtain the hue channel feature H3;

[0059] Step 7: The process of the adaptive block saturation enhancement function is as Figure 3 shown. Apply the adaptive block saturation enhancement function to the saturation channel feature S1 separated in Step 3 to obtain the enhanced saturation channel feature S2;

[0060] Step 8: Use channel fusion on the enhanced saturation channel feature S2 in Step 7 and the HSV feature L1 in Step 3 to obtain L3

[0061] Step 9: Apply a 3×3 convolution and a 5×5 convolution to L3 in Step 8 to obtain the saturation channel feature S3;

[0062] Step 10: Apply a 3×3 convolution to the saturation channel feature S3 in Step 9 to obtain the saturation channel feature S4;

[0063] Step 11: Use channel fusion on the value channel feature V1 separated in Step 3 and the HSV feature L1 to obtain L4;

[0064] Step 12: Apply a 3×3 convolution and a 5×5 convolution to L4 in Step 11 to obtain the value channel feature V2;

[0065] Step 13: Apply a 3×3 convolution to the value channel feature V2 in Step 12 to obtain the value channel feature V3;

[0066] Step 14: Based on a variant of the Retinex theory, perform a point-by-point multiplication operation on the value channel feature V3 obtained in Step 13 and the HSV feature L1 in Step 3 to obtain the value channel feature V4;

[0067] Step 15: Perform a pixel addition operation on the value channel feature V4 obtained in Step 14 and the HSV feature L1 in Step 3 to obtain the value channel feature V5;

[0068] The third part includes ten steps:

[0069] Step 16: For the fusion attention as Figure 4 shown, use the value channel feature V5 obtained in Step 15 and the saturation channel feature S4 obtained in Step 10 as inputs into the fusion attention to obtain the enhanced image F1;

[0070] Step 17: Apply convolution for channel expansion on the enhanced image F1 obtained in Step 16 and the hue channel feature H3 obtained in Step 6 to obtain the enhanced image F2 and the hue channel feature H4;

[0071] Step 18: As shown in the receptive field module RFB Figure 5 , taking the hue channel feature H4 in step 17 as the input, passing through the receptive field module RFB, and obtaining the enhanced image G1.1. The receptive field module RFB consists of a dual-branch attention, a residual depth separable convolution block RDSCB, and a normalization operation LN. The dual-branch attention is formed by connecting the guiding attention DA and the global attention GA in parallel;

[0072] Step 18-1: Taking the hue channel feature H4 in step 17 as the input, passing through the global attention GA, and obtaining the intermediate feature M1;

[0073] Step 18-2: Using the hue channel feature H2 in step 5 as the guide, taking the hue channel feature H4 in step 17 as the input, passing through the HAD module, and obtaining the intermediate feature M2, where HDA is the DA module guided by the hue channel feature H2 in step 5;

[0074] Step 18-3: Fusing the intermediate feature M1, the intermediate feature M2, and the hue channel feature H4 in step 17 to obtain the intermediate feature M3;

[0075] Step 18-4: Passing the intermediate feature M3 through the normalization operation LN and the residual depth separable convolution block RDSCB, and then fusing it with the intermediate feature M3 to obtain the output feature G1.1;

[0076] Step 19: Taking the output feature G1.1 of step 18 as the input of the receptive field module RFB, and obtaining the output feature G1.2;

[0077] Step 20: Taking the output feature G1.2 of step 19 as the input of the receptive field module RFB, and obtaining the output feature G1.3;

[0078] Step 21: Fusing the output feature G1.2 of step 19 and the output feature G1.3 of step 20, taking it as the input of the receptive field module RFB, and obtaining the output feature G1.4;

[0079] Step 22: Fusing the output feature G1.4 of step 21 and the output feature G1.1 of step 18, taking it as the input of the receptive field module RFB, and obtaining the output feature G1.5;

[0080] Step 23: Fusing the output feature G1.5 of step 22 and the hue channel feature H4 in step 17 to obtain the output feature G1.6;

[0081] Step 24: Take the enhanced image F2 in step 17 as input, take the saturation channel feature S3 in step 9 as a guide, repeat the operations from step 18 to step 23, and obtain the enhanced image G2.n, where n = 1, 2, 3, 4, 5, 6;

[0082] Step 25: Take the enhanced image F2 in step 17 as input, take the brightness channel feature V2 in step 12 as a guide, repeat the operations from step 18 to step 23, and obtain the enhanced image G3.n, where n = 1, 2, 3, 4, 5, 6;

[0083] The fourth part consists of a step:

[0084] Step 26: Fusion attention Figure 6 As shown, G1.6 obtained in step 23, G2.6 obtained in step 24, and G3.6 obtained in step 25 are put into the fusion attention as input to obtain the output image;

[0085] Part 5 consists of two steps:

[0086] Step 27: Input the training set samples in step 2 into the network from step 3 to step 26, set the network hyperparameters: learning rate is 2e-4, batch size is 4, optimizer is Adam, loss function is L1Loss, train the network to obtain the final interference image denoising pre-training model;

[0087] Step 28: Input the interference noise image test data set obtained in step 1 into the pre-trained model obtained in step 27, and the network can restore the interference image without noise or with little noise.

[0088] The present invention starts from the acquisition of interference data set and the problem of insufficient contrast and high noise of interference image, and provides an image denoising method for interference image. First, the method uses the interference formula and Legendre polynomials to generate interference simulation images, and randomly introduces various types of local block noise and overall noise into the interference simulation image. After artificial darkness adjustment, a paired interference image data set is obtained; secondly, in order to solve the problem of insufficient contrast of interference image, the method uses Retinex theory and introduces an adaptive saturation enhancement module to enhance different channels to solve the problem of insufficient contrast of interference image; then, in order to better deal with the inherent noise of interference image and the noise problem caused by brightness change, a three-branch denoising network is adopted, and a dual-branch attention mechanism is introduced to process images from multiple angles, so as to achieve better results; finally, by deploying two fusion attentions, a better feature integration effect is achieved. The algorithm of the present invention is simple, operable, and suitable for a variety of interference images.

[0089] Although the above description has been made of the illustrative specific embodiments of the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

Claims

1. An image denoising method for interference images, characterized in that, Image enhancement is performed using multi-channel processing, and dual-branch attention and fusion attention are used to denoise and fuse the enhanced images, including five parts: dataset preprocessing, channel feature enhancement and feature extraction, multi-angle image denoising, image fusion, and training and testing of the network model: The first part includes two steps: Step 1: Interference simulation images are generated by randomly selecting the order of the Legendre polynomial and randomly adjusting the phase coefficient, where the order is restricted between 15 and 400. The interference simulation images are used as normal light images, and the brightness is artificially adjusted to obtain low-light images. A variety of types of local block noise and overall noise are randomly introduced into the dataset to obtain interference noise images. A total of 1000 groups of interference image pairs are generated to form an interference noise dataset, which is used as training data and test data in a ratio of 4:1; Step 2: The interference image pairs are cropped into image blocks of the same size according to a sampling size of 128×128, and data augmentation operations are performed on the sampled data to expand the data samples to form the final training set samples; The second part includes thirteen steps; Step 3: The training sample images in Step 2 are input into the network, and the feature L1 in the RGB space is transformed into the HSV space through channel transformation, and three channels are separated, namely the hue channel feature H1, the saturation channel feature S1, and the value channel feature V1; Step 4: The hue channel feature H1 separated in Step 3 and the HSV feature L1 are fused through channel fusion to obtain L2; Step 5: L2 in Step 4 is passed through a 3×3 convolution and a 5×5 convolution to obtain the hue channel feature H2; Step 6: The hue channel feature H2 in Step 5 is passed through a 3×3 convolution to obtain the hue channel feature H3; Step 7: The saturation channel feature S1 separated in Step 3 is passed through an adaptive block saturation enhancement function to obtain the enhanced saturation channel feature S2; Step 8: The enhanced saturation channel feature S2 in Step 7 and the HSV feature L1 in Step 3 are fused through channel fusion to obtain L3; Step 9: L3 in Step 8 is passed through a 3×3 convolution and a 5×5 convolution to obtain the saturation channel feature S3; Step 10: The saturation channel feature S3 in Step 9 is passed through a 3×3 convolution to obtain the saturation channel feature S4; Step 11: The value channel feature V1 separated in Step 3 and the HSV feature L1 are fused through channel fusion to obtain L4; Step 12: L4 in Step 11 is passed through a 3×3 convolution and a 5×5 convolution to obtain the value channel feature V2; Step 13: The value channel feature V2 in Step 12 is passed through a 3×3 convolution to obtain the value channel feature V3; Step 14: Based on a variant of the Retinex theory, the value channel feature V3 obtained in Step 13 is multiplied point by point with the HSV feature L1 in Step 3 to obtain the value channel feature V4; Step 15: The value channel feature V4 obtained in Step 14 is added to the pixels of the HSV feature L1 in Step 3 to obtain the value channel feature V5; The third part includes ten steps: Step 16: Use the lightness channel feature V5 obtained in Step 15 and the saturation channel feature S4 obtained in Step 10 as inputs and put them into the fusion attention to obtain the enhanced image F1; Step 17: Use convolution to perform channel expansion on the enhanced image F1 obtained in Step 16 and the hue channel feature H3 obtained in Step 6 to obtain the enhanced image F2 and the hue channel feature H4; Step 18: Use the hue channel feature H4 in Step 17 as the input, pass it through the receptive field module RFB to obtain the enhanced image G1.

1. The receptive field module RFB is composed of a dual-branch attention, a residual depthwise separable convolutional block RDSCB, and a normalization operation LN. The dual-branch attention is formed by connecting the guiding attention DA and the global attention GA in parallel; (1) Use the hue channel feature H4 in Step 17 as the input, pass it through the global attention GA to obtain the intermediate feature M1; (2) Use the hue channel feature H2 in Step 5 as the guide, use the hue channel feature H4 in Step 17 as the input, pass it through the HAD module to obtain the intermediate feature M2, where HDA is the DA module guided by the hue channel feature H2 in Step 5; (3) Perform a fusion operation on the intermediate feature M1, the intermediate feature M2, and the hue channel feature H4 in Step 17 to obtain the intermediate feature M3; (4) Pass the intermediate feature M3 through the normalization operation LN and the residual depthwise separable convolutional block RDSCB, and then fuse it with the intermediate feature M3 to obtain the output feature G1.1; Step 19: Use the output feature G1.1 in Step 18 as the input of the receptive field module RFB to obtain the output feature G1.2; Step 20: Use the output feature G1.2 in Step 19 as the input of the receptive field module RFB to obtain the output feature G1.3; Step 21: Perform channel fusion on the output feature G1.2 in Step 19 and the output feature G1.3 in Step 20, use it as the input of the receptive field module RFB to obtain the output feature G1.4; Step 22: Perform channel fusion on the output feature G1.4 in Step 21 and the output feature G1.1 in Step 18, use it as the input of the receptive field module RFB to obtain the output feature G1.5; Step 23: Perform channel fusion on the output feature G1.5 in Step 22 and the hue channel feature H4 in Step 17 to obtain the output feature G1.6; Step 24: Use the enhanced image F2 in Step 17 as the input, use the saturation channel feature S3 in Step 9 as the guide, repeat the operations in Step 18 - Step 23 to obtain the enhanced image G2.n, where n = 1, 2, 3, 4, 5, 6; Step 25: Use the enhanced image F2 in Step 17 as the input, use the lightness channel feature V2 in Step 12 as the guide, repeat the operations in Step 18 - Step 23 to obtain the enhanced image G3.n, where n = 1, 2, 3, 4, 5, 6; The fourth part includes one step: Step 26: Use G1.6 obtained in Step 23, G2.6 obtained in Step 24, and G3.6 obtained in Step 25 as inputs and put them into the fusion attention to obtain the output image; The fifth part includes two steps: Step 27: Input the training set samples in Step 2 into the network from Step 3 to Step 26, and set the network hyperparameters: the learning rate is 2e-4, the batch size is 4, the optimizer is Adam, the loss function is L1Loss, and train the network to obtain the final interference image denoising pre-training model; Step 28: Input the interference noise image test dataset obtained in Step 1 into the pre-training model obtained in Step 27, and the network can recover the interference image without noise or with little noise.

2. The image denoising method for interference images according to claim 1, wherein, In the process of making the interference noise dataset in Step 1, the missing interference dataset in the public data is supplemented, and the dataset can be expanded at any time according to the needs.

3. A method for image denoising of interference images according to claim 1, characterized in that, The three branches in the second part and the third part are connected, and different processing is performed on the three channel features, which can enhance the image from multiple scales.

4. A method for image denoising of interference images according to claim 1, characterized in that The connection positions of the same fusion attention in Step 16 and Step 26 are different. Using the fusion attention at different connection positions can better retain information. The fusion attention in Step 16 fuses the two enhanced features, preventing over-enhancement and providing more detailed information for the denoising module. The fusion attention in Step 26 fuses the three denoised features, maximizing the retention of interference image information.

5. A method for image denoising of an interference image according to claim 1, characterized in that, The receptive field module RFB in Step 18 introduces dual-branch attention and processes the feature tensor hierarchically. The residual depth separable convolution block changes the connection of the depth separable convolution to a residual manner, further preventing the loss of details.

Images