Research On Real Noise Image Denoising Algorithm Based On Multi-scale Feature

As an important information transmission carrier in human daily life,the quality of image directly determines whether people can obtain effective information from it in timely and accurately.In the process of image acquisition,due to the imperfection of imaging system and equipment,as well as the influence of external factors such as illumination,digital image will be added different types of noise in the process of formation,transmission and storage,which will cause the loss of image information.With the rapid development of convolutional neural network in the field of image processing,the synthetic noise denoising algorithm based on convolutional neural network can not meet the needs of real-world noisy image denoising,and its practicability is insufficient.The complex network model can not meet the needs of efficient processing.Aiming at the above problems,the following three algorithms are studied in this paper:1.Aiming at the problems that a single feature can not effectively remove the complex real-world noise,the generalization ability and model robustness of real-world noisy image denoising algorithms need to be improved,a multi-scale residual denoising algorithm based on progressive training is proposed.Based on the multi-scale image obtained by average pooling,combined with multi-scale image information extraction and multi-scale feature extraction,an end-to-end denoising model is established.Densely connected residual blocks are designed to extract hierarchical features.The dilated convolution module can expand the receptive field and strengthen the mapping process between images.The hierarchical features of the upper layer are fed back to the next layer.This progressive training method can flexibly use the image features.The experimental results show that this algorithm can effectively deal with the denoising problem of real-world noisy images.2.Aiming at the problems that the image pyramid obtained by average pooling will bring a great amount of computation,and the insufficient fusion of multi-scale features is easy to cause information loss,a convolutional neural network denoising algorithm based on parallel multi-scale feature extraction is proposed.The algorithm takes the adaptive dense residual block as the horizontal network structure,and selectively enhances the characteristics of large amount of information in the channel.The vertical network structure uses the feature pyramid to realize the parallel extraction of multi-scale features.The features of the single horizontal scale are fully integrated with the features of different vertical scales,which enhances the role of local features on global features.This algorithm is more conducive to noise removal,and obtains a clearer and more natural restored image.3.Aiming at the problems that the middle scale of network layering is not fully utilized and can not adaptively enhance the features with large amount of information in image space,a real-world image denoising algorithm based on multi-scale feature reuse is proposed.On the basis of preserving adaptive dense residual blocks,a cross-scale feature fusion module is proposed to make full use of the intermediate features of each layer.The multi-scale spatial attention module is studied.Spatial attention is added after the dilated convolution with different dilation rates,and different computing resources are allocated for the image spatial information with different importance.The down sampling layering of the network structure is improved to further reduce the lack of information caused by the smaller scale.Experimental results show that this algorithm can restore images with clearer edges and texture details when removing more noise.