Research On Deep Convolutional Neural Networks On Image Restoration

Image restoration technique has become one of the important issues in the field computer vision.Due to environment,weather,camera and human factors,captured images are noisy and low-resolution in general.Thus,image denoising and image superresolution techniques of image restoration techniques have been widely concerned by researchers.Scholars have utilized signal processing,priori-based approaches and deep network techniques to propose a lot of solutions for image denoising and image supersolution in recent decades.However,most of these methods ignore the effect of both complex random noise and complex background on image denoising,both loss of highfrequency detailed information and complex low-resolution image(i.e.,low-resolution image of unknown corruption condition with and without noise)for image super-resolution.For addressing these problems,this paper combines the characteristic of the given task via signal processing and neural network knowledge to design efficient deep convolutional neural networks.Main contexts of this paper have the following parts:(1)To resolve image denoising problem of complex random noisy image denoising,this paper proposes a dual-path convolutional neural network for image denoising.This method extends deep network into wide network to obtain more complementary wide features for suppressing complex random noise.Using batch renormalization addresses internal covariate shift from samples in training processing on low-figuration platform.Using dilated convolutions extracts more features of deep layers in the network to enlarge the differences between different sub-networks for improving the denoising performance.Experimental results show that the proposed method can not only effectively address noisy image of known types,such as Gaussian noisy and real noisy images denoising problems,but also resolve uneven distributions of data on a low-configuration hardware platform during the training process for noisy images of known types denoising problem.(2)To solve image denoising in the complex backgrounds,this paper proposes an attentive convolutional neural network for image denoising.According to complexity and attributes of the network,this method designs an efficient denoising network via using attentive mechanism to extract noise information from complex backgrounds.To improve denoising performance and efficiency,using dilated convolutions and common convolutions implements a sparse mechanism in a CNN to enhance denoising effect.Taking into account long-term dependency problem from deep network,using long path fuses global and local features to enhance denoising effect of shallow layers on deep layers.Additionally,applying attentive mechanism via current stage to guide previous stage extracts salient noise from complex backgrounds.Finally,using residual learning technique can remove obtained noise from given noisy images.Quantitative analysis and qualitative analysis show that the proposed method can effectively deal with noisy images of both known types and unknown types(blind denoising)denoising problems.(3)To address loss of high-frequency detailed information for image super-resolution,this paper proposes a cascaded convolutional neural network for image super-resolution.This method fuses low-and high-frequency information of multiple different types to enhance training stability of super-resolution model.To address long-term dependency problem of the network,using heterogeneous convolutions obtains features of different types and fusing these features can increase effect of network shallow layers on deep layers.To prevent loss of edge information from repeated use of 1 × 1 convolutions in the heterogenous convolutions,applying residual learning technique fuses network hierarchical features to strengthen obtained features.To prevent unstable training from ignoring high-frequency detailed features caused by up-sampling operation of deep layer in the network,applying refinement block learns accurate high-frequency features and obtains a high-quality image.A lot of experiments prove that the proposed method can not only improve the stability of image super-resolution network,but also promote performance and efficiency for image super-resolution.(4)To address complex image super-resolution(low-resolution image restoration question of unknown corruption condition with and without noise),this paper presents an asymmetric convolutional neural network for image super-resolution.This method can reduce computational cost via enhancing effects of local salient features to reduce computational resource,and combine an adaptive sub-pixel convolutional technique to train complex image super-resolution model.Most of existing methods only simply fuse hierarchical features to promote performance of image super-resolution,however,it may increase convergence time of model in the training process and cause waste of computational resource.Thereby,this method uses one-dimensional asymmetric convolutions in vertical and horizontal directions to enhance the effects of square convolutional kernels for enlarging the importance of local salient features on image super-resolution task.To address long-term dependency problem in deep network,using residual learning technique gathers hierarchical low-frequency features obtained to enhance effects of shallow layers on deep layers in the network.Additionally,applying an adaptive sub-pixel convolutional technique can convert low-frequency features into high-frequency features to train certain scale image super-resolution model,varying scale image super-resolution model(blind super-resolution model)and varying scale image super-resolution model with noise of unknown types,which can make readers choose appropriate functions according to their own actual needs.Finally,using residual learning technique fuses high-and lowfrequency features to address unstable training problem from sampling operation of deep layer in the network.Experiment results show that the proposed asymmetric network has obtained superior performance in certain scale and complex image super-resolution(varying scale image super-resolution and varying scale with noise of unknown types image super-resolution).In summary,this paper starts from real needs by combining principles of network design,attributes of target task and challenges of application scenarios to present image restoration techniques for mobile devices,where their rationality and effectiveness via both method analysis and a lot of experiments are shown.