基于深度残差网络的单幅大倍率超分辨率重建模型研究

Super-resolution reconstruction is an effective technique to enhance image resolution and therefore has a wide range of practical applications. The current challenge of this area lies in the handling of single image super-resolution at large upscaling factors. Its difficulties are derived from both the loss of a large amount of high-frequency information and the space and scale mismatch between the images before and after reconstruction. Therefore, this project will focus on the research on single image super-resolution reconstruction model at large upscaling factors based on deep residual network. Firstly, to effectively alleviate the mismatch between the images before and after reconstruction, a post-sampling and feature-space-learning-based method is studied by the construction of a trainable stepwise upsampling mechanism and the usage of data manifold simplification, so as to improves the model prediction accuracy and reconstruction performance. Secondly, the redundancy of the BN structure in the network is verified in order to optimize the existing structure of the residual reconstruction network. Meanwhile, the convolutional layers of the network are upgraded and enlarged from the aspects of their depth, width and channel configuration. Based on this, an enhanced deep residual network is built, which will have larger size and improved performance. Finally, in order to break the limitation of the conventional loss function, a visual-perception-based loss function is designed. This loss considers the changes in the local structures and in each single pixel simultaneously. In doing so, it can improve the performance and robustness of the network after training. Through the above research, a novel method that can solve the problem of single image super-resolution reconstruction at large upscaling factors will be formed.

超分辨率重建是一种增强图像分辨率的有效技术，应用十分广泛。在该领域中，单幅图像的大倍率重建是现今难点，问题的困难在于大量高频信息的丢失以及重建前后图像空间尺度的极度不匹配。因此，本项目将围绕基于深度残差网络的单幅大倍率超分辨率重建模型展开研究。首先，通过构建可训练的逐步上采样机制同时配合数据流形简化思想，研究一种后置上采样的特征空间学习方法，有效缓解重建图像间的不匹配性，提升模型估计准确度及重建效果。其次，验证模型中BN结构的冗余性，以优化残差重建网络的现有结构，并从深度、宽度、通道配置等方面对网络中卷积成分进行升级拓展，构建一种增强的深度残差网络模型，提升网络规模及表现能力。最后，为打破传统损失函数的局限，研究设计基于视觉感知的损失函数，该损失同时考虑图像的局部结构变化和单个像素偏差，能够提升训练后网络的性能及鲁棒性。通过以上研究，将形成一种解决单幅大倍率超分辨率重建问题的新方法。

超分辨率重建是一种增强图像分辨率的有效技术，应用十分广泛。在该领域中，单幅图像的大倍率重建是现今难点，问题的困难在于大量高频信息的丢失以及重建前后图像空间尺度的极度不匹配。因此，本项目围绕基于深度残差网络的单幅大倍率超分辨率重建模型展开研究，主要研究进展包括：（1）构建了一种后置上采样的特征空间学习方法，有效缓解重建图像间的不匹配性，提升模型估计准确度及重建效果；（2）优化了残差重建网络的现有结构，据此提出了一种增强的深度残差网络模型，提升网络规模及表现能力；（3）设计了一种基于视觉感知损失函数的生成对抗式训练方法，打破传统损失函数的局限；同时，围绕项目相关扩展及应用问题：（4）提出了一种两段式卷积神经网络模型，有效解决零样本学习低照度图像增强中的色彩失真和噪声干扰问题；（5）探索了项目核心成果在林业特色领域中的应用，主要涵盖了工程竹材裂纹检测和林业遥感图像增强等两方面。本项目研究形成了一种解决单幅大倍率超分辨率重建问题的新方法。.整体来说，本项目完成了预定的研究内容，并在此基础上做出了进一步的扩展。在项目资助下，共计发表学术论文17篇，其中SCI期刊论文13篇，中文期刊论文4篇。申请国家发明专利5项，其中已授权1项。登记软件著作权2项。