基于自注意力深度网络的图像超分辨率重建方法

Image Super-Resolution Reconstruction Method Based on Self-Attention Deep Network

针对现有图像超分辨重建方法难以充分重建图像的细节信息且易出现重建的图像缺乏层次的问题,提出一种基于自注意力深度网络的图像超分辨重建方法。以深度神经网络为基础,通过提取低分辨率图像特征,建立低分辨率图像特征到高分辨率图像特征的非线性映射,重建高分辨率图像。在进行非线性映射时,引入自注意力机制,获取图像中全部像素间的依赖关系,利用图像的全局特征指导图像重建,增强图像层次。在训练深度神经网络时,使用图像像素级损失和感知损失作为损失函数,以强化网络对图像细节信息的重建能力。在3类数据集上的对比测试结果表明,所提方法能够提升图像超分辨重建结果的细节信息,且重建图像的视觉效果更好。

It is difficult to fully recover the image details using the existing image super-resolution reconstruction methods.Furthermore,the reconstructed images lack a hierarchy.To address these problems,an image super-resolution reconstruction method based on self-attention deep networks is proposed herein.This method,which is based on deep neural networks,reconstructs a high-resolution image using the features extracted from a corresponding low-resolution image.It nonlinearly maps the features of a low-resolution image to those of a high-resolution image.In the process of nonlinear mapping,the self-attention mechanism is utilized to obtain the dependence among all the pixels in the images,and the global features of the images are used to reconstruct the corresponding high-resolution image,which promotes image hierarchy.During the deep neural network training,a loss function comprising a pixel-wise loss and a perceptual loss is utilized to improve the image-detail reconstruction ability of the neural network.Experiments on three open datasets show that the proposed method outperforms the existing methods in terms of image-detail reconstruction.Furthermore,the visual impression of the reconstructed image is better than that of the images reconstructed using other existing methods.