Single Image Super Resolution Based On Deep Learning

Due to the limitations caused by both internal or external factors or conditions of the imaging system,the image which is obtained by the digital imaging system is of low resolution(LR)and few details,and difficult to meet application level.Single image super resolution(SISR)algorithm is supposed to be capable of recovering high resolution(HR)image from low resolution,so that to break out the limitations.The technology has been widely used in consumer electronics products,medical imaging,surveillance systems,remote sensing imaging and other fields.Existing super resolution algorithms can be classified into interpolation-based,reconstruction-based,and learning-based methods.Among them,the learning-based method has a better reconstruction quality than the other two,which has received extensive attention from researchers.Our study and research is stated in this paper concentrate on learning-based methods.The contents are summarized as follows.The neighbor embedding method uses the Euclidean distance to find neighbors when selecting neighborhoods.The similarity between the neighborhood and the input image being the largest is not guaranteed,the obtained reconstruction weight is not optimal,and the texture area of the reconstructed image is blurred.In order to address the problems above,perceptual hash based neighbor embedding single image super resolution algorithm is proposed.The algorithm utilizes perceptual hash to select the neighbors that are most similar to the input LR image patches.The k/K nearest neighbor selection algorithm is used to constrain the number of neighbor selections,and a HR image is finally synthesized by minimizing the reconstruction error.Experiments results analysis and comparison verified that the proposed algorithm is of high reconstruction quality.The existing deep learning based SISR algorithm uses a chain structure to build a network model.However,the extracted low level features are not used in the final image reconstruction process in a very deep convolutional network,hence the degraded reconstruction quality.In order to solve this problem,this paper proposes a single image super resolution algorithm based on deep learning via skip connection.This method is an end-to-end approach,with the interpolated LR image as input,using the skip connect to introduce the learned low level image features to the final reconstruction process.The learned low level image features by the former layer and the advanced image features obtained by the network end is merged with pixel-wise addition.Finally,HR image is reconstructed by residual learning.Experiments show that the algorithm proposed in this paper is able to effectively recover the missing texture in the image,and reconstructed quality is improved.The fact is that learning-based SISR has many limitations in the industrial applications on account of its high complexity and excessive hardware resources consumption.Aiming to apply the proposed technology in industry,we conduct an investigation on status of image super resolution technology applying in 4K TVs and smart phones,and introduce an industrial method to evaluate the quality of reconstructed images.Based on the above,the algorithm we proposed is tested on TV set and smart phones to verify its effectiveness.Finally,the feasibility of the proposed algorithm in hardware is analyzed.