针对目前基于样例学习的图像超分辨率方法难以同时满足快速运算和生成高质量图像的问题,提出一种基于去卷积的快速图像超分辨率方法。设计新型网络模型,以低分辨率图像作为输入图像,利用卷积层进行特征提取与表示;利用去卷积层对图像特征放大膨胀,再以池化层浓缩特征图,提炼出对结果更敏感的特征;以亚像素卷积层实现特征映射与图像融合,获得高分辨率图像。在图像集上进行测试,相比其他方法,本文方法的测试结果具有较高的峰值信噪比,且平均每秒能处理24幅以上大小为320 pixel×240 pixel的图像,表明该方法不仅可以生成更高质量的图像,且具有较高的处理速度,能满足视频实时处理要求。

In view of existing problems of image super-resolution method based on sample-learning, which is difficult to operate rapidly and generate high quality image at the same time, a rapid image super-resolution method based on deconvolution is proposed. A new type of network model is designed and low resolution images are taken as input images directly, and then convolution layer is used to extract and represent features. Deconvolution layer is used to enlarge image feature maps, and the following pooling layer is used to concentrate the feature maps and extract features which are more sensitive to the results. Moreover, sub-pixel convolution layer is applied to features mapping and images fusion simultaneously and the super-resolution image could be obtained. The proposed method is tested on images of test datasets, and compared with other methods. The test results of the proposed method have higher peak signal to noise ratio (PSNR) and can process more than 24 images in size of 320 pixel×240 pixel per second, which shows that the proposed rapid image super-resolution method based on deconvolution can not only generate images with higher quality, but also satisfy the requirement of real-time video processing.