Single Image Super Resolution Based On Deep Learning

Single image super resolution(SISR)aims at recovering high resolution images from a low resolution image by the mapping relationship between low resolution and high resolution images learnt in the training phase.Super resolution algorithms based on deep learning obtain better results in comparasion to traditional algorithms,but there still exits a gap between the perceptual quality of reconstruction images and human visual perception.This thesis digs into the problems exiting in recent algorithms,and then proposes our three new algorithms,which obtain high objective indices and better subjective results.(1)We propose a SISR algorithm based on Enhanced Inception Unit.For the problem that recent algorithms ignore the relationship of features extracted from adjacent convolutional layers,the proposed method builds the Enahced Inception Unit based on filter concatenation and residual connection to merge the features extracted from adjacent convolutional layers,and furthly it can compact the relationship of them.For the problem that middle and primary features are not effeciently used which makes the reconstruction quality poor,these units are connected by residual connections to form a local block,and it generates many blocks from the input to the ouput in the whole network.Finally,features extracted from every local block are merged in the construction layer by recursive way to improve reconstruction quality.The experiments indicate that this method obtain better image reconstruction quality than the state-of-art algorithms.(2)We propose a single image super resolution algorithm based on self attention,the algorithm contains holistic adversarial network based on self attention and local adversarial network based on local adversarial network.Because the local receptive field is limited for textures reconstruction,the method uses holistic adversarial network based on self attention to obtain the non-local information,and furtherly utilize more global information to reconstruct textures.SISR algorithms based on generative adversarial network can reconstruct images corresponding to human feelings,but the local textures may not be realistic,the method propose local adversarial network based on image patches,which judges the reconstructed image patches and original image patches are real or fake,this make the local textures of reconstrction image become more natural.Extensive experiments indicate the method can recover realistic local texture,as well as obtain the non-local information,which make the reconstruted images have more similarities with original images.(3)We propose a SISR algorithm based on textures selection and structure preservation.Most SISR algorithms only consider the single degradation distortion,however,real scenario distorted images contain many distortions,so,the results are not desired for these images.To address these issues,the method reconstructs images based on the characteristics of real scenario distortion images.To recover textures,the method selects texture features from shallow convolutional layers by lateral modules.To recover structures,the method adopts the mix of L1 and SSIM loss functions to make the images have more sharpness.To improve the reconstruction quality furtherly,we introduce smooth dilated convolution to enlarge the receptive field.Extensive experiments indicate that our algorithm can recover textures and structures effectively based on standard datasets and real scenario distorted images.