4K low-illumination video denoising method based on multiple stages

By building a multi-stage self-supervised denoising network model, combining blind domain and local perception network modules, adding attention mechanism and model pruning, the problems of high noise and computational complexity in low-illumination video denoising are solved, and the problem of high noise and computational complexity in low-illumination video denoising is realized. Efficient noise removal effect on edge devices.

CN118982482AInactive Publication Date: 2024-11-19SICHUAN NATIONAL INNOVATION VISION UHD VIDEO TECHNOLOGY CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411053343.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-02
Publication Date
2024-11-19
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The prior art has high image noise and reduced contrast in low illumination environments, and the calculation of denoising algorithms is costly and complex, making it difficult to effectively apply in practical applications.

Method used

Using a multi-stage 4K low-illumination video denoising method, a self-supervised denoising network model is constructed, including a blind domain network module, a local perception network module and a denoising network module, a convolutional attention module is added, and a self-supervised training and modeling is used. Pruning, reducing computational complexity and model size.

Benefits of technology

Achieving stable, fast and accurate denoising functions on devices with limited computing resources improves algorithm efficiency and convenience and improves noise denoising effect.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118982482A_ABST
    Figure CN118982482A_ABST
Patent Text Reader

Abstract

The invention discloses a 4K low-illumination video denoising method based on multiple stages. The method comprises the following steps: constructing and training a self-supervised denoising network model; inputting a to-be-denoised low-illumination video into the self-supervised denoising network model to obtain a denoised low-illumination video; the self-supervised denoising network model comprises a blind domain network module, a local sensing network module and a denoising network module; the blind domain network module learns flat region noise features and obtains a first denoising result; the local sensing network module learns non-flat noise features and obtains a second denoising result; and the de-noising network module obtains a de-noised low-illumination video according to the first de-noising result and the second de-noising result. The method has the beneficial effects that the self-supervised denoising network model is constructed to effectively reduce the influence of noise on image restoration, so that the method has better local texture restoration capability, reserves details and has better anti-noise performance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a multi-stage 4K low-illumination video denoising method. Background Art

[0002] One of the problems facing digital high-definition images now is that the image noise is high and the contrast is reduced in low-light environments, the effective recognition rate of the image is low, and the noise in the low-light images after image enhancement is often still not eliminated. Therefore, it is becoming increasingly important to further denoise the enhanced low-light images in order to actively and effectively improve the image effect.

[0003] In recent years, the use of convolutional neural networks to construct denoising algorithms has become the mainstream method in the field of image denoising and video denoising. However, the computational cost of these algorithms is too high and the network structure is very complex, which limits their application in practice. It is very necessary to design lightweight denoising convolutional neural networks for practical applications and solve the problems of computational cost and complexity. The commonly used method for lightweighting convolutional neural networks is to reduce network parameters and floating-point operations (FLOPs), such as using smaller convolutions and reducing the number of convolutions, but this method often cannot guarantee algorithm performance.

[0004] To address this deficiency, a multi-stage 4K low-illumination video denoising method is proposed. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a more lightweight and multi-stage 4K low-light video denoising method with better noise reduction effect, so that it can provide stable, accurate and fast denoising function on edge devices with limited computing resources and low power consumption, thereby improving the efficiency and convenience of the algorithm.

[0006] A multi-stage 4K low-light video denoising method is characterized by comprising:

[0007] Build and train a self-supervised denoising network model;

[0008] Input the low-light video to be denoised into the self-supervised denoising network model to obtain the denoised low-light video;

[0009] Among them, the self-supervised denoising network model includes a blind area network module, a local perception network module and a denoising network module;

[0010] The blind area network module is used to denoise the input 4K low-light video to obtain a first denoising result;

[0011] The local perception network module is used to denoise the input 4K low-light video to obtain a second denoising result;

[0012] The denoising network module is used to denoise the 4K low-light video according to the first denoising result and the second denoising result to obtain a denoised video.

[0013] Furthermore, the blind domain network module and the local domain network module are both provided with a convolutional attention module, and the convolutional attention module includes a channel attention module and a spatial attention module;

[0014] The channel attention module is used to obtain a channel attention feature map according to the input feature map;

[0015] The spatial attention module is used to obtain a spatial attention feature map according to the input feature map;

[0016] The input feature map is a preliminary feature map obtained after the 4K low-light video enters the blind area network module and the local area network module respectively.

[0017] Furthermore, the channel attention mechanism is expressed as follows:

[0018] M C (F)=sigmoid(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0019] Where:

[0020] F: input feature map;

[0021] AvgPool: average pooling;

[0022] MaxPool: maximum value pooling;

[0023] Mc(F): Channel attention feature map.

[0024] Furthermore, the expression of the spatial attention module is as follows:

[0025] M S (F) = sigmoid(f 7×7 ([AvgPool(F); MaxPool(F)]))

[0026] Where:

[0027] F: input feature map;

[0028] AvgPool: average pooling;

[0029] MaxPool: maximum value pooling;

[0030] Ms(F): spatial attention feature map;

[0031] f 7×7 (x) represents a convolution layer with a convolution kernel size of 7×7, and the channel of the output spatial attention feature map is 1.

[0032] Furthermore, building and training a self-supervised denoising network model includes:

[0033] Construct a self-supervised denoising network model, which includes a blind area network module, a local perception network module and a denoising network module;

[0034] Acquire a training data set, where the training data set includes a real noise image; the real noise image is acquired based on a 4K low-light video;

[0035] Obtaining noise characteristics of a noisy image;

[0036] Input the noisy image into the blind area network module to obtain the denoised image of the flat area;

[0037] Obtaining a flat area noise feature according to the flat area denoising image, and obtaining a first denoising result through the flat area noise feature and the noise feature of the noise image;

[0038] Input the noisy image into the local area network module to obtain the denoised image in the non-flat area;

[0039] Acquire the non-flat region noise feature according to the non-flat region denoising image, and obtain the second denoising result according to the first denoising result, the non-flat region noise feature and the noise feature of the noise image;

[0040] A self-supervised training method is used to train a denoising network module according to the first denoising result and the second denoising result to obtain a pre-trained model;

[0041] The clear image and Gaussian noise are input into the pre-trained model, and the clear image and Gaussian noise are mixed to obtain a synthetic image; the synthetic image corresponds to the clear image;

[0042] Using the self-supervised training method, the pre-trained model is trained based on the clear image and the synthetic image to obtain the final self-supervised denoising network model;

[0043] The expression of the denoising loss function when training the pre-trained model based on clear images and synthetic images is:

[0044] The expression of the denoising loss function is as follows:

[0045] L d =||G(x+noise)-x||1;

[0046] Where:

[0047] x: clear image;

[0048] noise: Gaussian noise.

[0049] Furthermore, the expression of the first loss function of the blind field network module is as follows:

[0050]

[0051] Where:

[0052] The flat area denoised image at the output of the blind field network module;

[0053] y: Noise image input in the training set;

[0054] The first loss function can ensure that the global denoising result processed by the blind field network module, that is, the flat area denoised image, is consistent with the characteristic structure of the original noisy image.

[0055] Furthermore, the expression of the second loss function of the local perception network module is as follows:

[0056]

[0057] in,

[0058] 0<α<1, α is a rational number;

[0059] A first denoising result obtained from the flat region noise feature map;

[0060] Denoising images in non-flat areas.

[0061] Furthermore, the expression of the third loss function of the denoising network module is as follows:

[0062]

[0063] Where:

[0064] A first denoising result obtained by denoising the image in the flat area;

[0065] a second denoising result obtained according to the non-flat region denoising image;

[0066] Comprehensive denoising results.

[0067] Further, adopting a self-supervised training method, training a denoising network module according to the first denoising result and the second denoising result to obtain a pre-trained model includes:

[0068] Use L1 norm to perform sparse training on the normalization layer in the denoising network module to obtain the weight parameter file;

[0069] S8001, use L1 norm to perform sparse training on the batch normalization layer in the denoising network module to obtain the weight parameter file; the sparse factor sr is 0.005,

[0070] S8002. Sort the importance of weights by global weight importance according to the weight parameters;

[0071] S8003, obtaining the channel ordering between convolutional layers in the denoising network module according to the global weight importance ordering, the scaling factor γ and the bias β;

[0072] S8004. Obtain a new denoising network module according to the channel sorting.

[0073] Furthermore, the constraint calculation expression of the scaling factor is as follows:

[0074] g(γ)=|γ|;

[0075] Where:

[0076] g(γ): L1 regularization term with respect to γ.

[0077] The present invention has the following advantages:

[0078] (1) The present invention adds an attention module to the blind field network module and the local field network module, thereby improving the ability of the blind field network module to distinguish flat areas, and improving the ability of the local field network module to extract features of textures and non-flat areas, further enhancing the learning efficiency of the network module, and laying a foundation for improving the ability of the denoising network module;

[0079] (2) Only low-light video is input to obtain the original image. Using the original image as the self-supervised denoising network module during pre-training will result in poor denoising results in the denoising network module, because there is no reverse constraint process in the entire learning stage to guide all networks to compare and learn the features of clear images. The present invention inputs clear images and Gaussian noise, adds branches to strengthen self-supervision - simultaneously inputs a synthetic image of Gaussian synthetic noise and clear images, and then compares and analyzes the denoising output results with the clear data in reverse, further guiding the stage three denoising network to learn the features of clear data.

[0080] (3) The present invention performs model pruning and lightweighting on the denoising network, reducing the model size and computational complexity, so that it can run on resource-constrained devices such as edge devices. The denoising capability is improved by the above-mentioned innovation of adding an attention module and inputting clear images and Gaussian noise to ensure the algorithm capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 It is a diagram of the method steps of the present invention;

[0082] Figure 2 The comparative experiment verification data diagram of the present invention. DETAILED DESCRIPTION

[0083] The present invention is further described below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following description.

[0084] It should be noted that the directions or positional relationships indicated by "left", "right", etc. are based on the directions or positional relationships shown in the accompanying drawings, or are the directions or positional relationships in which the inventive product is usually placed when in use, or are the directions or positional relationships commonly understood by those skilled in the art. Such terms are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention.

[0085] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features and technical solutions in the embodiments may be combined with each other.

[0086] like Figure 1 As shown, the present invention proposes a multi-stage 4K low-illumination video denoising method, comprising:

[0087] Build and train a self-supervised denoising network model;

[0088] Input the 4K low-light video to be denoised into the self-supervised denoising network model to obtain the denoised 4K low-light video;

[0089] Among them, the self-supervised denoising network model includes a blind area network module, a local perception network module and a denoising network module;

[0090] The blind area network module is used to denoise the input 4K low-light video to obtain a first denoising result;

[0091] The local perception network module is used to denoise the input 4K low-light video to obtain a second denoising result;

[0092] The denoising network module is used to denoise the 4K low-light video according to the first denoising result and the second denoising result to obtain a denoised video.

[0093] In the first stage, the method uses a blind field network module to learn the flat area noise of the input 4K low-light video to obtain a first denoising result. In the second stage, the local perception network module uses a local perception network module to learn the non-flat noise of the input 4K low-light video to obtain a second denoising result.

[0094] The feature learning of the denoising network module in the third stage is trained based on the first denoising results and the second denoising results to obtain a comprehensive denoising result. When denoising the actual video, only the denoising network module in the third stage can be used for denoising. Therefore, the high-order information learning ability of the first two stages can be enhanced, which can empower the lightweight third-stage denoising network module and further improve the algorithm running speed while ensuring accuracy.

[0095] Furthermore, the blind domain network module and the local domain network module are both provided with a convolutional attention module, and the convolutional attention module includes a channel attention module and a spatial attention module;

[0096] The channel attention module is used to obtain a channel attention feature map according to the input feature map;

[0097] The spatial attention module is used to obtain a spatial attention feature map according to the input feature map;

[0098] The input feature map is a preliminary feature map obtained after the 4K low-light video enters the blind area network module and the local area network module respectively.

[0099] Furthermore, the expression of the channel attention module is as follows:

[0100] M C (F)=sigmoid(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0101] Where F represents the input feature map;

[0102] AvgPool means average pooling;

[0103] MaxPool means maximum pooling;

[0104] Mc(F) represents the channel attention feature map.

[0105] Furthermore, the expression of the spatial attention module is as follows:

[0106] M S (F) = sigmoid(f 7×7 ([AvgPool(F); MaxPool(F)]))

[0107] Where F represents the input feature map;

[0108] AvgPool means average pooling;

[0109] MaxPool means maximum pooling;

[0110] Ms(F) represents the spatial attention feature map;

[0111] f 7×7 (x) represents a convolution layer with a convolution kernel size of 7×7, and the channel of the output spatial attention feature map is 1.

[0112] Furthermore, building and training a self-supervised denoising network model includes:

[0113] S1. Construct a self-supervised denoising network model, which includes a blind area network module, a local perception network module and a denoising network module;

[0114] S2. Obtain a training data set, where the training data set includes a real noise image; the real noise image is obtained based on a 4K low-light video;

[0115] S3, obtaining noise characteristics of the noise image;

[0116] S4, inputting the noisy image into the blind area network module to obtain a denoised image of the flat area;

[0117] S5. Obtaining a flat area noise feature according to the flat area denoised image, and obtaining a first denoising result through the flat area noise feature and the noise feature of the noise image;

[0118] S6, inputting the noisy image into the local area network module to obtain a denoised image in the non-flat area;

[0119] S7, obtaining the non-flat region noise feature according to the non-flat region denoising image, and obtaining the second denoising result according to the first denoising result, the non-non-flat region noise feature and the noise feature of the noise image;

[0120] The first denoising result is used to supervise the learning of the local perception network module in the second stage and the denoising module in the third stage;

[0121] S8, using a self-supervised training method to train a denoising network module according to the first denoising result and the second denoising result to obtain a pre-trained model;

[0122] S9, inputting the clear image and Gaussian noise into the pre-trained model, mixing the clear image and the Gaussian noise to obtain a synthetic image; the synthetic image corresponds to the clear image;

[0123] S10, using a self-supervised training method to train a pre-trained model based on the clear image and the synthetic image to obtain a final self-supervised denoising network model;

[0124] The expression of the denoising loss function when training the pre-trained model based on clear images and synthetic images is:

[0125] L d =||G(x+noise)-x||1;

[0126] in,

[0127] x: clear image;

[0128] noise: Gaussian noise.

[0129] Furthermore, the expression of the first loss function of the blind field network module is as follows:

[0130]

[0131] in, It is the flat area noise feature map output by the blind field network module;

[0132] y is the input noisy image.

[0133] A blind neighborhood network module is used to remove spatially correlated noise in flat areas. The blind neighborhood network module here is an improvement on the blind spot self-supervised network BSN. BSN only deducts the single pixel in the center, while the blind neighborhood network module deducts a larger area in the middle and extracts pixels farther away from the center point, which is reasonable for flat areas and can better reduce the influence of surrounding noise pixels. The loss function of the blind neighborhood network module can ensure that the global denoising result processed by the blind neighborhood network module is consistent with the characteristic structure of the original noisy image.

[0134] Furthermore, the loss function of the local perception network module is expressed as follows:

[0135]

[0136] Among them, 0<α<1, α is a rational number;

[0137] A first denoising result obtained from a flat region denoising image;

[0138] Denoising images in non-flat areas.

[0139] The local perception network module uses the flat area denoised image obtained by the blind neighborhood network module as supervision information for training, and the trained local perception network module is used in the texture area to generate texture details;

[0140] The blind neighborhood network module in the previous stage processes the entire original image obtained from the 4K low-light video to generate That is, the feature image of the flat area; after the image is processed by the blind neighborhood network module, the local variance of the flat area is small, and the local variance of the texture area is large, so the α area can be generated. The platform area in this area is low and the flat area can be extracted, while the 1-α area will highlight the flat area, which is not conducive to extracting the flat area;

[0141] The loss function of the local perception network module ensures that the outputs of the blind neighborhood network module and the flat area of ​​the local perception network module are consistent after local perception by the local perception network module, which is used to distinguish between texture and flat areas.

[0142] Furthermore, the loss function of the denoising network module is expressed as follows:

[0143]

[0144] A first denoising result obtained by denoising the image in the flat area;

[0145] a second denoising result obtained according to the non-flat region denoising image;

[0146] Comprehensive denoising results.

[0147] The loss function of the denoising network module is used to fuse the denoising results of the blind neighborhood network module and the local perception network module to ensure that the denoising effects of the flat area and the texture area are different.

[0148] Further, adopting a self-supervised training method, training a denoising network module according to the first denoising result and the second denoising result to obtain a pre-trained model includes:

[0149] S8001, using L1 norm to perform sparse training on the batch normalization layer in the denoising network module to obtain a weight parameter file; in some specific embodiments, the sparse factor sr used in the sparse training is 0.005;

[0150] S8002. Sort the importance of the weights according to the weight parameters;

[0151] S8003, obtaining the channel ordering between convolutional layers in the denoising network module according to the global weight importance ordering, the scaling factor γ and the bias β;

[0152] S8004. Obtain a new denoising network module according to the channel sorting.

[0153] Furthermore, the constraint calculation expression of the scaling factor γ is as follows:

[0154] g(γ)=|γ|;

[0155] Where:

[0156] g(γ): L1 regularization term with respect to γ.

[0157] The channels represented by the scaling factor γ close to 0 after sparsification have a low contribution rate to the denoising network module. Through this training, the importance of the channels of the convolutional layer can be distinguished, and these channels and the input and output of the convolutional layer can be pruned to remove redundant channels. This can achieve the effect of compressing the denoising network module structure and improving the computing speed without affecting the performance of the denoising network module.

[0158] In this embodiment, an attention module and a channel pruning algorithm are added to the self-supervised denoising network model, so that the self-supervised denoising network model can quickly denoise the noisy image of 4K high-resolution low-light video and ensure the stability of denoising.

[0159] A comparative experiment was constructed, using the peak signal-to-noise ratio (PSNR) and structural similarity ratio (SSIM), which are commonly used in the image field, as indicators for evaluating denoising results. The larger the PSNR, the higher the image quality, and the higher the SSIM, the more similar the original image is to the denoised image.

[0160] The comparative experiments include the evaluation indicators of GID, CBM3D1, CBDNET, Spatially Adaptive SSID models on DnD, SIDD, PolyU, and HighISO datasets; the experimental results are as follows: Figure 2 ;

[0161] Figure 2 In the figure, when the first row method is GID, the PSNR is 34.32 and the SSIM is 0.817 when the data set is the DnD data set. The higher the PSNR data, the higher the PSNR image quality. The higher the SSIM, the closer it is to the clear data. It is believed that the denoising result is closer to the clear image. Therefore, Figure 2 It can be seen that when the solution of the present invention is applied to various types of training data, the time used by the present invention is the shortest when the PSNR and SSIM data are relatively high, that is, when the image is of high quality.

[0162] The above embodiments only express preferred implementation modes, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention.

Claims

1. A multi-stage 4K low-light video denoising method, characterized by: include: Build and train a self-supervised denoising network model; Input the 4K low-light video to be denoised into the self-supervised denoising network model to obtain the denoised 4K low-light video; Among them, the self-supervised denoising network model includes a blind area network module, a local perception network module and a denoising network module; The blind area network module is used to denoise the input 4K low-light video to obtain a first denoising result; The local perception network module is used to denoise the input 4K low-light video to obtain a second denoising result; The denoising network module is used to denoise the 4K low-light video according to the first denoising result and the second denoising result to obtain a denoised video.

2. The multi-stage 4K low-light video denoising method according to claim 1, characterized in that: The blind domain network module and the local domain network module are both provided with a convolutional attention module, and the convolutional attention module includes a channel attention module and a spatial attention module; The channel attention module is used to obtain a channel attention feature map according to the input feature map; The spatial attention module is used to obtain a spatial attention feature map according to the input feature map; The input feature map is a preliminary feature map obtained after the 4K low-light video enters the blind area network module and the local area network module respectively.

3. The multi-stage 4K low-light video denoising method according to claim 2, characterized in that: The expression of the channel attention module is as follows: M C (F)=sigmoid(MLP(AvgPool(F))+MLP(MaxPool(F))) Where: F: input feature map; AvgPool: average pooling; MaxPool: maximum value pooling; Mc(F): Channel attention feature map.

4. The multi-stage 4K low-light video denoising method according to claim 2, characterized in that: The expression of the spatial attention module is as follows: M S (F)=sigmoid(f 7×7 ([AvgPool(F);MaxPool(F)])) Where: F: input feature map; AvgPool: average pooling; MaxPool: maximum value pooling; Ms(F): spatial attention feature map; f 7×7 (x) represents a convolution layer with a convolution kernel size of 7×7, and the channel of the output spatial attention feature map is 1.

5. The multi-stage 4K low-light video denoising method according to claim 1, characterized in that: Building and training a self-supervised denoising network model includes: Construct a self-supervised denoising network model, which includes a blind area network module, a local perception network module and a denoising network module; Acquire a training data set, where the training data set includes a real noise image; the real noise image is acquired based on a 4K low-light video; Obtaining noise characteristics of a noisy image; Input the noisy image into the blind area network module to obtain the denoised image of the flat area; Obtaining a flat area noise feature according to the flat area denoising image, and obtaining a first denoising result through the flat area noise feature and the noise feature of the noise image; Input the noisy image into the local area network module to obtain the denoised image in the non-flat area; Acquire the non-flat region noise feature according to the non-flat region denoising image, and obtain the second denoising result according to the first denoising result, the non-flat region noise feature and the noise feature of the noise image; A self-supervised training method is used to train a denoising network module according to the first denoising result and the second denoising result to obtain a pre-trained model; The clear image and Gaussian noise are input into the pre-trained model, and the clear image and Gaussian noise are mixed to obtain a synthetic image; the synthetic image corresponds to the clear image; Using the self-supervised training method, the pre-trained model is trained based on the clear image and the synthetic image to obtain the final self-supervised denoising network model; The expression of the denoising loss function when training the pre-trained model based on clear images and synthetic images is: L d =||G(x+noise)-x||1; Where: x is a clear image; noise is Gaussian noise.

6. The multi-stage 4K low-light video denoising method according to claim 5, characterized in that: The loss function of the blind field network module is expressed as follows: Where: The flat area denoised image at the output of the blind field network module; y: Noise image of the training set input.

7. The multi-stage 4K low-light video denoising method according to claim 5, characterized in that: The expression of the loss function of the local perception network module is as follows: Where: 0<α<1, α is a rational number; A first denoising result obtained from a flat region denoising image; Denoising images in non-flat areas.

8. The multi-stage 4K low-light video denoising method according to claim 5, characterized in that: The loss function of the denoising network module is expressed as follows: Where: A first denoising result obtained by denoising the image in the flat area; a second denoising result obtained according to the non-flat region denoising image; Comprehensive denoising results.

9. The multi-stage 4K low-light video denoising method according to claim 8, characterized in that: Using a self-supervised training method, training a denoising network module according to the first denoising result and the second denoising result to obtain a pre-trained model includes: S8001. Use L1 norm to perform sparse training on the batch normalization layer in the denoising network module to obtain a weight parameter file; S8002. Sort the importance of weights by global weight importance according to the weight parameters; S8003, obtaining the channel ordering between convolutional layers in the denoising network module according to the global weight importance ordering, the scaling factor γ and the bias β; S8004. Obtain a new denoising network module according to the channel sorting.

10. The multi-stage 4K low-light video denoising method according to claim 9, characterized in that: The constraint calculation expression of the scaling factor γ is as follows: g(γ)=|γ|; Where: g(γ): L1 regularization term with respect to γ.

Citation Information

Patent Citations

  • Fine-grained image classification method based on sparse bilinear convolutional neural network

    CN110147834A

  • Fire identification and early warning method and system

    CN111882810A

  • Edge-guided single image noise removal

    CN115631107A

  • Image denoising method based on spatial adaptive self-supervised learning

    CN116071265A