Research On Image Restoration Based On 3D Spatio-temporal Convolution

Image restoration technologies mainly include image denoising,image enhancement,image super-resolution,etc.They have been widely used in various fields such as aerospace technology,remote sensing technology,medical imaging,and social communication.In the field of image denoising,it is further divided into Stripe noise,Gaussian noise,Poisson noise and so on according to the characteristics of noise distribution.The existence of image noise has seriously affected the accurate acquisition of image spatial information by machine vision systems,and has brought great challenges to subsequent tasks such as image recognition,image classification,and image detection and segmentation.Therefore,it is important to study image noise and restore it.This paper first studies the traditional three-dimensional block matching(BM3D),non-local mean(NLM),wavelet transform image denoising methods.Most of them are based on the principle of frequency domain or space domain,which can remove noise to a certain extent,but the traditional method is too complicated and inefficient to solve the optimization problem.The non-convex denoising model is basically used,and the model parameters cannot be automatically updated.Scene adaptability is poor.Compared with traditional image denoising methods,image denoising methods based on deep neural networks can better solve the above problems,but existing deep neural network models still have difficulty in achieving balanced results in terms of removing noise and maintaining details.In view of this,this paper presents a stripe noise deep neural network model that utilizes the space-time features of sequence images.Aiming at the invariance of stripe noise,this model builds a spatio-temporal information extraction module,and extracts the spatio-temporal correlation feature of adjacent frame images through a 3D convolution kernel to better estimate the noise.In view of the 3D convolution and the 2D convolution,the amount of parameters increases multiples,so the 2D convolution is used in the reconstruction module to improve the calculation efficiency.In addition,the algorithm also introduces multi-scale convolution to extract multi-scale spatial features,and designs a multi-residual reconstruction network to further improve the accuracy of image denoising.Experiments show that the model is superior to current mainstream image denoising algorithms in both subjective and objective indicators.In order to further restore the image details and reduce the calculation time,this paper proposes an algorithm model based on wavelet convolutional neural network.The model fully considers the inherent characteristics of strip noise and the complementary information between different wavelet sub-band coefficients,so that the noise can be accurately estimated with a low amount of calculation.For different subbands,this paper designs different spatiotemporal feature extraction subnetworks,and then fuses the features extracted from each subband to reconstruct the denoised image.The input of the entire network is a multi-frame sequence image after wavelet decomposition,and the output is a single-frame denoised image.Experiments show that the algorithm can greatly improve the denoising performance of the network without increasing the amount of calculation.