针对单一先验知识不足以约束病态严重的图像超分辨率问题, 本文提出了融合内外特征的图像超分辨率算法。针对图像的自相似性, 通过采用基于内部特征的深度卷积网络学习来增强输入图像的细节纹理, 去除超分辨率图像伪影; 同时, 使用基于外部图像的稀疏约束方法来学习图像结构信息, 并结合高频残差字典来解决超分辨率重建中的高频信息缺失问题; 最后通过卷积稀疏方法分别从基础层和细节层来融合内外特征的重建图像, 以获得细节清晰、去伪影的超分辨率图像, 进一步提高图像质量。与传统算法相比, 本文算法在重建图像的纹理特征和质量上都得到了增强, 且视觉效果与峰值信噪比较传统算法有所改善。

Existing knowledge on the severely ill-posed super-resolution problem is not sufficient; therefore, we propose an algorithm that joint constraints the internal and external features of image. For the self-similarity of the image, a deep convolutional network based on the internal features is used to enhance the detailed texture of the input image and remove the artifacts of the super-resolution image. Further, we utilize the sparse constraint method based on the external features to obtain the structural information of the image, and the high-frequency residual dictionary is combined to solve the problem of loss of high-frequency information in the reconstruction of super-resolution images. Finally, the features of the basic and detail layers in the reconstruction image are fused by the convolutional sparse method to obtain the super-resolution image with clear details and without artifacts, further improving the image quality. Compared with the traditional algorithms, the proposed method enhances the texture feature and quality of the reconstructed image. It also considerably improves the visual quality and peak signal to noise ratio.