Reseach On Image Super-Resolution Via Semantic Segmentation

In many domains of digital images,people often desired for High Resolution images.Although image sensors have been widely used to capture digital images,with the improvement of resolution requirements,current electronic components are gradually unable to meet the future demand in terms of price and function.Therefore,it is necessary to find approaches to enhance the resolution of digital images.Image super-resolution is a very method that can enhance the resolution of digital images.Image super-resolution is a long-standing and challenging research in the field of image processing and computer vision,which is aiming to estimate or restore a high resolution image from its low resolution counterpart.As different high resolution images may generate the same low resolution images by downsampling,image super-resolution in essence is an typical ill-posed and one-to-many problem.Image super-resolution has received extensive attention from within the computer vision research community and has a wide range of applications in consideration of its challenging difficulty and high application value.The research of computer vision has been divided into high and low two levels.The high-level task tends to image comprehension,mainly containing image recognition,semantic segmentation and object detection;while the low-level task tends to image processing,mainly containing denoising,deblur,inpainting,super-resolution and recovering raw images from compressed images.In recent years,despite the breakthroughs in accuracy and speed of image super-resolution using deep neural networks,essence of image super-resolution is merely still identified as the low-level task,which is deficient in neglecting the information in high level that can reinforce and constrain the result in image processing.Based on the problem above,this paper combine the semantic segmentation and image super-resolution and proposed an image super-resolution framework based on semantic segmentation.The image super-resolution framework is composed of semantic segmentation networks and cascaded reconstruction networks in series,semantic segmentation networks extract semantic feature map and semantic layout of the image to be reconstructed,and cascaded reconstruction networks use semantic layout and semantic feature map for image super-resolution.In addition,by studying and adjusting the weight of the loss in the perceptual loss,the cascaded reconstruction networks have more visual persuasion and sense of reality in the reconstructed results.This paper explored the effect of semantic segmentation networks on cascaded reconstruction networks through experiments and the effect of loss function on super-resolved results.After the mean opinion score testing,the average gains on mean opinion score achieved by our method are higher than the next appraoch SRGAN are 0.16 for the upscaling factor 4,which comfirms that our method is more visual suasive and realistic than the several reference algorithms for big upscaling factor.