Image Super-Resolution Based On Deep Information-Compensation Network

Computer vision supports the development of artificial intelligence.As an important carrier of information,people have higher and higher requirements for the quality of image.However,due to the distance between the camera and the object,the limitation of lighting conditions,the camera shake during shooting,and the poor quality of the hardware equipment,the captured image may be small or in a blurry state,thus producing a low-resolution image.In order to obtain high-quality images,the goal of the image super-resolution task is to recover the corresponding high-resolution images from the low-resolution images with severe information loss.In recent years,deep convolution neural networks have learned high-level semantic feature representations through multi-layer non-linear structures.Although they perform well under constrained conditions,they are limited by the lack of accurate structured information priors,and existing methods are difficult to obtain accurate texture structure information.To deal with this problems,the main researches of this paper are as follows:1.Considering that traditional face super-resolution reconstruction methods lack prior knowledge of face geometry,we propose a geometric compensation networks approach for interactive face alignment and image super-resolution,which typically reconstructs high-resolution faces and detects facial landmarks iteratively on the given low-quality facial images.Our method exploits the explicit interactions between face super-resolution and landmark detection under a unified deep learning paradigm.To achieve this,we learn to localize facial landmark coordinates based on the synthetic high-resolution images each time.Simultaneously,our model reasons out the attentional mask with these coordinates and recovers facial details on the low-resolution images.The experimental results validate that this method can effectively use the prior knowledge of face geometry,and the peak signal-to-noise ratio on the face data set 300W is improved by about 7%.2.Traditional perceptual super-resolution reconstruction methods do not make full use of the image self-similarity.This paper proposes an attention compensation network method for natural image perceptual image super-resolution reconstruction.This method is dedicated to learning the optimal perceptual loss function and searching for the optimal spatially adaptive attention super-resolution network framework in a unified reinforcement learning framework.Specifically,this paper designs a learnable perceptual super-resolution loss function in order to effectively obtain image perception information.In addition,in order to further improve the perceptual quality of the image,this paper designs an attention operation in the search space,which learns the edge structured information of the object in a non-local way.Experimental results show that this method can make full use of the image self-similarity to extract the structured information of the edge of the object,and the learned perceptual image patch similarity(LPIPS)on the natural image data set Urban100 with complex structure is reduced by about 14%.3.According to the above research,the image super-resolution reconstruction system based on the deep information-compensation network is designed and implemented.The system takes low-resolution face images or low-resolution natural images as input,and uses the network model proposed in this paper to complete the image super-resolution reconstruction task.The practicability of the system is fully verified by testing the samples in the face dataset and natural image dataset.