Search

CN-116433521-B - CBDNet denoising network optimization method and system for enhancing edge characteristics

CN116433521BCN 116433521 BCN116433521 BCN 116433521BCN-116433521-B

Abstract

The invention discloses a CBDNet denoising network optimization method and system for enhancing edge characteristics, wherein the method comprises the steps of obtaining a real image dataset and a synthetic image dataset, alternately inputting an original image into a noise estimation network, outputting a noise estimation diagram, splicing the original image and the noise estimation diagram, inputting the spliced image into a non-blind denoising network, outputting a denoised image, and adopting a U-Net architecture for the non-blind denoising network, wherein the non-blind denoising network comprises a plurality of convolution layers, two bilinear interpolation upsampling layers and two IndexNet modules. According to the CBDNet denoising network optimization method and system for enhancing the edge characteristics, the effect of subsequent image processing or detection is improved by reducing the loss of image edge information in the denoising process, more image texture characteristics are reserved by using a CBDNet non-blind denoising network mode, and meanwhile, the influence caused by excessive denoising due to overestimated noise level of a noise estimation network is reduced.

Inventors

  • SHAN YUXIANG
  • GAO YANGHUA
  • YU GANG

Assignees

  • 浙江中烟工业有限责任公司

Dates

Publication Date
20260512
Application Date
20230412

Claims (9)

  1. 1. A CBDNet denoising network optimization method for enhancing edge features, comprising: Respectively acquiring a real image data set and a synthetic image data set; Alternately inputting the original images selected from the real image data set and the synthetic image data set into a noise estimation network, and outputting a noise estimation graph; Splicing the original image and the noise estimation graph; The method comprises the steps of inputting the spliced original image and the noise estimation image into a non-blind denoising network, outputting the denoised image, wherein the non-blind denoising network adopts a U-Net architecture and comprises a plurality of convolution layers, two bilinear interpolation upsampling layers and two IndexNet modules, wherein the non-blind denoising network comprises 5 convolution blocks, 1 convolution layer, 2 averaging pooling layers, two bilinear interpolation upsampling layers and two IndexNet modules, the first convolution block consists of 2 convolution layers, the second convolution block consists of 3 convolution layers, the third convolution block consists of 6 convolution layers, the fourth convolution block consists of 3 convolution layers, the fifth convolution block consists of 2 convolution layers, the first IndexNet module is added after the first convolution block and is used for guiding the first averaging pooling layer and the second bilinear interpolation upsampling layer, and the second IndexNet module is added after the second convolution block and is used for guiding the second averaging pooling layer and the first bilinear interpolation upsampling layer.
  2. 2. The method for optimizing a CBDNet de-noising network for enhanced edge features according to claim 1, wherein said separately obtaining a real image dataset and a composite image dataset specifically comprises: shooting by adopting real imaging equipment to obtain the real image data set; acquiring a noiseless image data set, adding poisson-Gaussian noise into each noiseless image in the noiseless image data set, and then performing demosaicing and gamma correction processes on the noiseless image added with the poisson-Gaussian noise to obtain the synthesized image data set.
  3. 3. The method for optimizing a CBDNet de-noising network for enhanced edge features according to claim 1, wherein said alternately inputting the original image selected from the real image dataset and the synthetic image dataset into a noise estimation network and outputting a noise estimation map specifically includes: The original image input into the noise estimation network has the same size as the noise estimation image output by the noise estimation network, and the noise estimation network adopts a full convolution network without a pooling layer and a BN layer.
  4. 4. The method of claim 3, wherein the original image input to the noise estimation network and the noise estimation image output by the noise estimation network each have a resolution of H×W×3, the noise estimation network comprises 5 ReLU-activated standard convolution layers, the first 4 convolution layers comprise 32 convolution kernels, the last convolution layer comprises 3 convolution kernels, and the convolution kernels of all the convolution layers are 3×3.
  5. 5. The method for optimizing a CBDNet de-noising network for enhancing edge features according to claim 4, wherein said stitching said original image and said noise estimate map specifically comprises: and splicing the original image and the noise estimation graph into a vector with the resolution of H multiplied by W multiplied by 6 by adopting a concatate function.
  6. 6. The method of claim 1, wherein each of the IndexNet modules includes 1 IndexBlock, sigmoid activation function and Softmax activation function, wherein IndexBlock is comprised of 1 convolutional layer and 1 pixel reconstruction layer as learning modules, Inputting the spliced original image and the noise estimation image into a non-blind denoising network, and outputting a denoised image, wherein the method specifically comprises the following steps of: The spliced original image and the noise estimation image are sent to IndexBlock, the length and width of the image are halved through a convolution layer with the convolution kernel step length of 2 and the output channel number of 4, and then the four feature images are combined into a feature image through a pixel reconstruction layer, and the size of the feature image is restored to the original size; The image output by IndexBlock is used for guiding the up-sampling process after being subjected to a Sigmoid activation function, the image output by IndexBlock and the image before up-sampling are subjected to element multiplication, and then up-sampling operation is carried out; The image obtained by the image IndexBlock through the Sigmoid function and then through the Softmax activation function is used for guiding the downsampling process, the image IndexBlock is subjected to element multiplication with the image before downsampling, downsampling operation is performed, and each element in the image is multiplied by 4 after downsampling.
  7. 7. The method for optimizing CBDNet denoising network for enhancing edge features according to claim 6, wherein said inputting the spliced original image and the noise estimate map into a non-blind denoising network, outputting a denoised image, specifically comprises: And inputting the spliced original image with the resolution of H multiplied by W multiplied by 6 and the noise estimation image into a non-blind denoising network, gradually reducing pixels after 2 times of IndexNet-guided average pooling layers, and recovering the pixels to the original size after 2 times of IndexNet-guided bilinear interpolation upsampling layers, thereby outputting an image with the resolution of H multiplied by W multiplied by 3.
  8. 8. The method of claim 6, wherein the noise estimation network and the non-blind denoising network are trained according to a total loss function, and parameter updating is performed by adopting a random gradient descent method, and the noise estimation network and the non-blind denoising network are trained together, wherein the total loss function is formed by weighting an asymmetric loss function and a total variation loss function of an output position of the noise estimation network, and a pixel-level mean square error loss function of the output position of the non-blind denoising network, and Calculating the asymmetric loss function of the noise estimation network output location by the following formula , (1) Wherein, the Representing an image input to the noise estimation network, Representing the estimated value of the intensity of the noise, Representing noise intensity of group-Truth noise-free image for , When the result is true, the data is displayed, Otherwise By assigning values To apply a greater penalty to underestimating noise strength; Calculating the total variation loss function of the noise estimation network output position by the following formula , (2) Wherein, the A gradient in the horizontal direction is indicated, Representing the gradient in the vertical direction; Representation of A norm; Calculating a pixel-level mean square error loss function of the output position of the non-blind denoising network according to the following formula , (3) Wherein, the The result of the denoising is indicated, Representing a group-Truth noise-free image, The total loss function is calculated by the following formula, (4) Wherein, the Representing an asymmetric loss function Is used for the weight of the (c), Representing a total variation loss function And not used when training with real data Consider only And At this time 。
  9. 9. A CBDNet denoising network optimization system that enhances edge features, comprising: the data set acquisition module is used for respectively acquiring a real image data set and a synthetic image data set; The noise estimation module is used for alternately inputting the original images selected from the real image data set and the synthetic image data set into a noise estimation network and outputting a noise estimation graph; the image stitching module is used for stitching the original image and the noise estimation image; The non-blind denoising module is used for inputting the spliced original image and the noise estimation image into a non-blind denoising network and outputting the denoised image, wherein the non-blind denoising network adopts a U-Net architecture and comprises a plurality of convolution layers, two bilinear interpolation upsampling layers and two IndexNet modules, the non-blind denoising network comprises 5 convolution blocks, 1 convolution layer, 2 average pooling layers, two bilinear interpolation upsampling layers and two IndexNet modules, the first convolution block consists of 2 convolution layers, the second convolution block consists of 3 convolution layers, the third convolution block consists of 6 convolution layers, the fourth convolution block consists of 3 convolution layers, the fifth convolution block consists of 2 convolution layers, and the first IndexNet module is added after the first convolution block and is used for guiding the first average pooling layer and the second bilinear interpolation upsampling layer, and the second IndexNet module is added after the second convolution block and is used for guiding the second average pooling layer and the first bilinear interpolation upsampling layer.

Description

CBDNet denoising network optimization method and system for enhancing edge characteristics Technical Field The invention relates to the technical field of image processing, in particular to CBDNet denoising network optimization method and system for enhancing edge characteristics. Background Image denoising is an important means of digital image processing, aimed at reducing various noise introduced or artificially added by the imaging process. Denoising algorithms are often used as an important element of the image preprocessing stage, or to enhance image sharpness. Today, denoising techniques seek to eliminate as much noise as possible and ensure that the information in the image is not corrupted by excessive denoising. The existing image denoising algorithm has good denoising effect. If CBDNet adopts a serial branch structure, the noise level is estimated through the noise estimation network, and then the denoising operation is performed according to the noise level through the non-blind denoising network. However, although CBDNet has a good denoising effect, detail loss may be caused by the pooling process, and a semantic gap problem exists in U-Net, so that image edges are blurred. Therefore, what is needed is a CBDNet denoising network optimization method and system that enhances edge features. Disclosure of Invention The invention aims to provide a CBDNet denoising network optimization method and system for enhancing edge characteristics, which are used for solving the problems in the prior art, and can optimize a non-blind denoising network model in CBDNet and keep edge details as much as possible. The invention provides a CBDNet denoising network optimization method for enhancing edge characteristics, which comprises the following steps: Respectively acquiring a real image data set and a synthetic image data set; Alternately inputting the original images selected from the real image data set and the synthetic image data set into a noise estimation network, and outputting a noise estimation graph; Splicing the original image and the noise estimation graph; And inputting the spliced original image and the noise estimation image into a non-blind denoising network, and outputting the denoised image, wherein the non-blind denoising network adopts a U-Net architecture and comprises a plurality of convolution layers, two bilinear interpolation up-sampling layers and two IndexNet modules. The method for optimizing CBDNet denoising network for enhancing edge features as described above, wherein preferably, the acquiring real image dataset and the composite image dataset respectively specifically includes: shooting by adopting real imaging equipment to obtain the real image data set; acquiring a noiseless image data set, adding poisson-Gaussian noise into each noiseless image in the noiseless image data set, and then performing demosaicing and gamma correction processes on the noiseless image added with the poisson-Gaussian noise to obtain the synthesized image data set. The method for optimizing CBDNet denoising network for enhancing edge features as described above, wherein preferably, the alternately inputting the original image selected from the real image dataset and the synthetic image dataset into the noise estimation network, outputting a noise estimation graph specifically includes: The original image input into the noise estimation network has the same size as the noise estimation image output by the noise estimation network, and the noise estimation network adopts a full convolution network without a pooling layer and a BN layer. The CBDNet denoising network optimization method for enhancing the edge characteristics is characterized in that the original image input into the noise estimation network and the noise estimation image output by the noise estimation network are all preferably H multiplied by W multiplied by 3 in resolution, the noise estimation network comprises 5 standard convolution layers activated by ReLU, the first 4 convolution layers comprise 32 convolution kernels, the last convolution layer comprises 3 convolution kernels, and the convolution kernels of all the convolution layers are all 3 multiplied by 3 in size. The method for optimizing CBDNet denoising network for enhancing edge features as described above, wherein preferably, the splicing the original image and the noise estimation graph specifically includes: and splicing the original image and the noise estimation graph into a vector with the resolution of H multiplied by W multiplied by 6 by adopting a concatate function. The CBDNet denoising network optimization method for enhancing the edge characteristics as described above, wherein preferably, the non-blind denoising network comprises 5 convolution blocks, 1 convolution layer, 2 average pooling layers, two bilinear interpolation upsampling layers and two IndexNet modules, wherein the first convolution block consists of 2 convolution layers, the second convolution blo