Application Research Of Camera Image And Video Denoising Algorithm Based On Convolutional Neural Network

In real life,cameras are widely used in video monitoring and data collection,such as license plate recognition in parking lots.However,under the condition of imperfect hardware or insufficient light,the image taken by the camera will generate more noise.If the 3D digital noise reduction that comes with the camera image processor is directly used,it can not achieve good denoising effect.On the other hand,with the development of deep learning technology,more and more computer vision tasks(such as face recognition,etc.)are applied in the realtime video surveillance systems.The accuracy of these visual tasks is closely related to the quality of the image or video collected by the camera.Therefore,the denoising of camera image or video is a problem worthy of further study.2D and 3D Digital Noise Reduction(DNR)are two traditional image and video denoising algorithms used by cameras.These two algorithms combine various filters in the spatial and time domains for denoising.However,there are many kinds and combinations of filtering,which require constant parameter adjustment to achieve the expected denoising effect,resulting in a relatively low degree of automation in denoising.In order to solve this problem,this thesis proposes a CNN-CameraDenoise(CNN-based Camera Denoise Framework)for the camera system combined with Convolutional Neural Network(CNN)in deep learning,and proposes an image denoising model based on CNN,resunetplus VGg,for the collected data set containing real noise.(1)CNN-CameraDenoise frameworkThe CNN-CameraDenoise framework includes CNN-based single-frame image denoising,multi-frame image denoising and video denoising.The CNN-based image denoising model can self-learn feature information in image data in a data-driven manner,which greatly reduces the workload of complex manual parameter adjustment required by traditional denoising algorithms.The simulation experiments of the CNN-CameraDenoise framework verify the feasibility of the framework.After the image denoising and video denoising of the CNNCameraDenoise framework,the visual effects and objective evaluation indicators of the images and videos are superior to the 3D noise reduction result of the camera interior.Moreover,the video picture after denoising is more stable and clear.(2)ResUNetplus-VGG modelBased on the UNet++network,the ResUNetplus-VGG model uses the pre-trained VGG network as a loss network based on transfer learning,so that the ResUNetplus-VGG model can better extract image details.At the same time,combining residual learning with the UNet++network,a better denoising effect can be obtained through the fusion of shallow and deep features.The experimental results show that the visual effect of the image after denoising by the ResUNetplus-VGG model has been significantly improved,and the objective evaluation indicators PSNR and SSIM values are better than other comparative models.