Super-resolution Reconstruction Of Face Images Based On Attention Mechanism And Prior Informatio

With the increasing demand for image quality,it is insufficient to improve the resolution of images only at the hardware level,and people have begun to focus on improving the resolution of images at the software algorithm level.Image super-resolution reconstruction technology refers to the use of some image processing algorithms and techniques to reconstruct one or more low-resolution images into high-resolution images.In recent years,with the unremitting efforts of many scholars,many excellent image super-resolution reconstruction algorithms have emerged,among which the most eye-catching is the image super-resolution reconstruction method based on deep learning.However,there are few super-resolution algorithms for face images currently.Existing reconstruction networks cannot distinguish primary and secondary information well.Under high magnification factors,problems such as image distortion and blurred facial features often occur.Face images in unrestricted scenes are reconstructed with super-resolution.This paper takes the face image as the research object.The face image super-resolution reconstruction network is constructed,and the attention mechanism is introduced to the network.Besides,residual learning,image semantic segmentation and other ideas are introduced into the reconstruction network.Finally,the network structure is optimized and improves the loss function.The main work of this paper is as follows:1)A super-resolution reconstruction method for face images based on hybrid attention mechanism and residual learning is proposed.By establishing a hybrid attention mechanism model in the channel domain and the spatial domain,the ability of the network to represent primary and secondary information is improved.Combining the residual learning theory to learn the high-frequency information of the image,add short jump connections to the attention mechanism model to construct the attention residual unit.At the same time,recursive learning is used to deepen the network layer and expand the experience field.The experimental results show that the network can reconstruct a clearer and high-quality face image under a variety of magnification factors,and the peak signal-to-noise ratio and structural similarity far exceed the image super-resolution method based on interpolation.2)For the super-resolution reconstruction of face images in unrestricted scenes,the prior information of face images is added to the network using image semantic segmentation algorithms to segment faces and backgrounds,and filter out useless backgrounds to strengthen facial features.At the same time,a generative confrontation network is introduced,and clear facial texture details are generated through the continuous game between the generation network and the confrontation network.The experiments show that under the 8x magnification factor,and the contours of the generated face images are clear.Besides,the texture details are also improved,and the visual effect is significantly improved.3)In order to further improve the reconstruction results under high magnification factors,we optimize the structure of the up-sampling network,and improve the traditional back-end up-sampling network into a progressive up-sampling network.The up-sampling method adopts sub-pixel convolution.The peak signal-to-noise ratio of the reconstruction result of the8 x magnification factor is improved by 0.16 d B.At the same time,perceptual loss and confrontation loss are introduced into the loss function to further increase the facial details of the reconstruction result.