Research On Image Denoising Algorithm Based On Convolutional Neural Network

Images are an important source of information for humans,and they have a more intuitive effect than words.However,in the process of acquiring and transmitting,images are susceptible to interference from electronic devices and light,resulting in noise and affecting people’s acquisition and recognition of information.Therefore,it is particularly important to study image denoising,which is also an important research content in the field of computer vision.In recent years,deep learning has been widely used because of its flexible network architecture and powerful learning ability of data models.In particular,convolutional neural networks in deep learning have achieved good performance in many image processing tasks.Based on convolutional neural networks,this paper proposes a network model based on channel attention for Gaussian denoising tasks in images,and the main work content is as follows:1.Based on channel attention,an efficient denoising network is designed in this paper,which can also meet the blind denoising task.The network structure can be divided into five modules,including sparse block,feature fusion block,feature compression block,information exchange block and image reconstruction block.Sparse blocks obtain global and local features by alternating between extended convolution and ordinary convolution.Feature fusion block fuses the global features extracted from the previous layers of networks to provide supplementary information for the next layer of networks.The feature compression block refines the extracted information and compresses the network to reduce the dimension.Information interaction block is used for feature integration and dimension reduction.The image reconstruction block subtracts the input noise image from the noise information learned by the network through residual learning,and finally gets the denoised image.2.In order to improve the denoising performance and reduce the learning difficulty of the network,this paper embedded the channel attention module into the network,and combined the batch normalization with residual learning.In order to improve the convergence speed of network training,L1 loss is combined with L2 loss when constructing loss function.3.In order to solve the problem of error accumulation and gradient disappearance caused by small mini-batch in the deep network training,the batch re-normalization technology is used in the improved network,so that the network can get better results with a small training sample,and it is applied in the denoising of CT images.4.In order to verify the performance of the proposed algorithm,the proposed algorithm is compared with other classical algorithms and algorithms based on deep learning.Experimental results show that the proposed algorithm can obtain good results under different noise levels.In conclusion,the proposed algorithm has excellent performance in image denoising.