基于深度学习的快速图像超分辨率方法

A Method for Fast Image Super-resolution Based on Deep Learning

图像超分辨率技术通过软件处理的方式,将输入的低分辨率的图片转化为相应的高分辨率图片,同时预测缺失的细节信息。针对现有的图像超分辨率模型重构效果较差、计算量较大等缺点,提出一种基于深度学习框架的快速超分辨率方法。模型的输入采用原始尺寸的低分辨率图片,大大减少了网络计算量。在特征提取阶段,采用循环卷积提取输入图像的特征信息;在图像重构阶段,采用并行的1×1卷积层对提取到的特征进行降维,并通过亚像素卷积得到相应的高分辨率图像。实验结果表明,相比现有的算法,提出的算法在超分辨率重构效果上更佳,且满足实时重构的要求。

Image super-resolution technology transforms the low-resolution images into corresponding high-resolution images and predicts the missing details via software processing.In the view of the shortcomings of existing super-resolution models,such as worse reconsitution result,large amount of computation,a fast super-resolution method based on deep learning is proposed in this paper.Original low-resolution images are used as the inputs of the proposed model,which significantly reduces the network calculation amount.In the feature extraction stage,features of the input low-resolution images are extracted by recursive convolution layers.For image reconstruction,parallelized 1×1 convolution layers are used to reduce the dimension of the extracted features,and corresponding high resolution image is gained through sub-pixel convolution.Experimental results show that the super-resolution reconstruction effect of proposed algorithm outperforms that of existing methods,while meeting the request of real-time reconstitution.