Research On Multi-scale Deep Learning Methods For Image Restoration

In recent years,image restoration methods based on deep learning have gradually become a mainstream study.However,there is still much room for such methods to improve when processing image details sensitive to the human visual system.An important reason for this problem is the lack of effective guidance in the existing methods.For that,this paper introduces traditional multi-scale image presentation techniques,as a kind of guidance,into popular deep learning techniques and therefore conducts research on multi-scale deep learning methods for image restoration.To achieve the goal mentioned above,a basic multi-scale deep learning framework for image restoration is first established based on a systematic review of existing multiscale methods.The fundamental idea is to give a deep learning network(called the "seed network" in this paper)scale property and then use multi-scale networks to process an input image simultaneously.With the established framework,a number of specific researches are conducted,which can be summarized as follows:1)A multi-scale deep learning algorithm based on image pyramid representation is proposed.It first performs multi-scale enhancement on the training data using image pyramid representation for training the "seed network" on multiple scales.Then,the trained multi-scale networks are simultaneously used to process the input image,followed by producing the final image restoration result through multi-scale fusion,achieving an overall performance improvement.2)A multi-scale deep learning algorithm based on image subband representation is proposed.It utilizes the advantage of image subband representation(i.e.,achieving a multi-scale decomposition of different image components in the frequency domain)to effectively guide the"seed network" for processing an input image at different scales and,therefore,achieves a high-quality image restoration result.3)A multi-scale deep learning algorithm based on image content representation is proposed.It begins with a spatial-wise multi-scale image content representation technique,which is designed based on the difference curvature feature,to decompose various contents of the input image into multiple scales.Then,a "seed network" is equipped to each scale for conducting differentiated processing on different image contents,yielding a high-quality image restoration result.4)Research on multi-scale feature extraction techniques inside a network is also conducted in this paper.The designed techniques include an attention-aware multi-receptivefield feature extraction module,an asynchronous multi-receptive-field feature extraction module,a spatial-wise multi-scale feature extraction module and a channel-wise multi-scale feature extraction module.With these techniques,a number of stable network models are further developed.To verify the advance and effectiveness of the proposed methods mentioned above,a large number of experiments for image restoration(including image interpolation,image super-resolution and image artifacts removal)have been carried out in this paper for algorithm analysis and evaluation.The experimental results have consistently shown the superiorities of the implemented algorithms on related tasks.