基于块旋转和稀疏表示的图像超分辨率重建

Image Super Resolution Based on Patch Rotation and Sparse Representation

摘要: 在智能车应用领域，高分辨率的图像已经成为汽车功能模块中不可或缺的一部分。然而传统的基于稀疏表示的高分辨率图像重建方法中所用的训练样本块特征单一，这就导致需要大量的样本块来训练字典。为了减少训练样本块，本文提出一种基于块旋转策略和稀疏表示的超分辨率重建算法。通过将图像块旋转不同的角度，从而减少样本块，增加特征数量，丰富训练字典的类型。在重建过程中，采用自适应加权求和的方式求得高分辨率图像。实验证明，所提出的方法较传统的方法，不仅在主观质量上有明显的提升，在客观质量上也有较大幅度的提高。

Abstract: In the intelligent vehicle applications, high resolution image has become an integral part of auto-mobile function module. However, the feature of the training image patch of the traditional sparse representation based reconstruction method is unitary, which leads to a large number of sample patches to train a dictionary. In order to reduce the number of training samples, this paper pro-posed a method based on patch rotation and sparse representation. By rotating the patch for dif-ferent angle, the number of the patches is reduced, the feature of the patch is increased and the dictionary type becomes rich. During the reconstruction process, the adaptive weighted method is used to obtain the high resolution image. Experiments show that, the proposed method compared with the traditional method, not only has significant improvement in subjective quality, but also greatly improves the objective quality.