Researches On Texture And Perception Guided Image Super-Resolution

Image super-resolution is a technique for improving image resolution by reconstructing a high-resolution image with more details from a low-resolution blurred image.As a key technology of image processing in computer vision,image super-resolution has been widely concerned by researchers.At the same time,image super-resolution is useful in semantic segmentation,pattern identification,and other domains,as well as improving people’s visual experience.Image super-resolution methods based on deep learning have gradually become a research hotspot in the field of image super-resolution due to the high quality of the generated images.Although great progress has been made in deep learning-based image super-resolution algorithms,there are still many challenging problems.For example,many algorithms pay too much attention to indicators such as peak signal-to-noise ratio and ignore the visual perceptual quality of the generated image.The methods based on generative adversarial networks can improve image perceptual quality,but indicators such as peak signal-to-noise ratio are reduced,and there is texture blurring.In this thesis,we conduct an in-depth analysis of the challenging problems of image super-resolution in perceptual quality and improving the reconstruction effect of textured regions.Cominbing with gradient information,visual masking mechanism,attention mechanism and generative adversarial network,we carry out the study on image super-resolution algorithms.The main contributions of this thesis are as follows.1.A texture and attention guided image super-resolution algorithm is proposed.In this thesis,a generative adversarial network is used to build an image super-resolution model.In order to alleviate the problem of poor details in the generated images of the image super-resolution algorithm based on generative adversarial networks,the proposed method improves the image quality by enhancing the detailed texture area,and it is based on directional gradient information.The loss function guides the model to focus on textured regions in the image during training.At the same time,in order to make the generated images more in line with human visual perception,the attention mechanism is used to build an attention image reconstruction network in the proposed method,so that the method can model the channel relationship of the up-sampling features.Thereby,the sensitivity of the model to the feature channel is improved,and the performance of image super-resolution is effectively improved.2.An image super-resolution algorithm combining visual masking mechanism and attention mechanism is proposed.On the basis of the former work,the multi-directional gradient information is used to improve the reconstruction quality of image super-resolution texture area further.At the same time,both the attention mechanism and the visual masking mechanism are considered.For the first time,the Just Noticeable Difference(JND)is introduced into the research of image super-resolution.By combining with the attention mechanism,the network that is more in line with the visual perception characteristics is developed,so that the model can achieve comprehensive optimization of the visual sensitive area and visual attention area,and further improve the visual quality of the reconstructed image.