资源受限环境下实时超分辨重建方法研究

Example learning has been recognized as an effective way to produce a high-resolution (HR) image from a single low-resolution (LR) input. However, due to the limitations of storage and computational capacity, many popular example learning-based super-resolution (SR) algorithms suffer from complicate learning structures and highly intensive computation, which makes them difficult to be applied in many practical applications. As a result, the key to successful applications of SR technique is to develop a fast yet effective algorithm that can be applied in most resource-limited scenarios. To this end, this proposal is going to make research on developing a fast single image SR approach that incorporates the advantages of both the example learning-based and reconstruction-based SR methods. In this proposal, a structured incoherence dictionary learning algorithm is proposed to represent multiple linearly coupled feature subspaces of LR and HR images. By establishing multiple linear mapping functions from the LR to HR images, a fast example regression-based SR algorithm is developed, which enables to maintain significantly low time and space complexity while making no compromise on quality. Furthermore, based on sum of squared difference (SSD), we propose a fast SR enhancement algorithm for improving the quality of SR images obtained by the example regression-based approach, wherein a unified regularization term that assembles local structure regularity and non-local similarity is integrated into the reconstruction-based SR framework for optimization. To fairly evaluate the quality of reconstructed results and the advantages and weaknesses of different SR methods, a no-reference image quality assessment metric for super-resolved images, which is consistent with the subjective perception, is going to be investigated, intending to provide a reliable evidence for the optimization of SR algorithms and the choice of related parameters.

实例学习是一种有效提高图像分辨率的超分辨重建技术。然而，受存储空间和计算能力的限制，很多实例学习超分辨算法由于学习结构复杂，计算密集度高，在实际中难以推广应用。因此，研究计算资源受限环境下的实时性超分辨重建算法，是该技术成功应用的关键。本课题拟研究结合基于实例学习和基于重构方法的快速单帧图像超分辨重建技术。提出一种基于"结构非相关"字典学习算法，获取表征低分辨与高分辨图像结构的多个线性特征子空间，通过学习低分辨与高分辨图像间多个线性映射关系，设计计算成本低、重建质量高的快速实例回归超分辨重建算法；同时，提出一种利用差值平方积分图技术的快速超分辨图像增强算法，将局部结构正则化、非局部相似性融合成一个统一的正则项，结合到基于重构超分辨框架下优化求解。为合理评判重建图像质量，进而评价重建算法的优劣，研究与人类主观感受相一致的无参考型图像质量评价准则，为超分辨重建算法的优化与参数选择提供可靠依据。

超分辨重建是一种有效提高图像分辨率的信号处理技术，在计算机视觉、医学与遥感成像和生活娱乐等具有广泛的应用前景。尽管人们提出了很多有效的超分辨重建算法，但大多数算法学习结构复杂，时间和空间复杂度均较高，很难在计算资源受限的环境下应用。为有效解决上述问题，本课题利用统计机器学习技术，研究了基于多个线性映射关系的快速实例回归超分辨重建方法、基于图像结构正则化的超分辨图像增强方法和无参考型的超分辨图像质量评价方法，在超分辨重建及其质量评价方面取得了一定的研究成果。. 图像超分辨方面，提出一种基于多线性映射关系学习的快速超分辨重建方法，降低了学习结构的复杂度和计算的密集度；提出了一种基于由粗到精的单针超分辨重建方法，有效提高了实例学习超分辨重建图像的质量；提出了一种基于相似性约束的结构化输出回归图像超分辨重建方法，能获得比多个单回归方法更好的重建质量；提出了一种基于主动采样和高斯回归的单针图像超分辨方法，相比于传统的高斯回归方法，提高了模型训练的效率和重建图像质量；提出了一种基于局部高斯过程的图像超分辨方法，有利于提升算法的性能；提出了一种Student-t似然下基于高斯过程回归的单帧图像超分辨方法，缓解了大规模数据集算法复杂度过高而限制应用的问题；提出了一种基于非局部可控核回归正则化的超分辨重建方法、基于局部字典学习和相似性结构正则化的超分辨重建方法，有利于重建边缘清晰的高质量图像。. 超分辨图像质量评价方面，提出了两层级联回归的无参考型超分辨图像质量评价方法，该方法从局部频域、全局频域和空域提取感知统计特征量化超分辨图像的失真程度，然后集成adaboost决策树回归和脊回归预测超分辨图像的质量，该方法与人眼视觉系统主观评价结果具有较好一致性。. 培养博士后1名，博士生1名，硕士生5名；发表与课题相关并标注本基金资助的学术论文共计17篇，其中SCI检索11篇，包括《IEEE Trans》5 篇。申请国家发明专利授权6项，获省部级奖励1项。