面向图像复原问题的可学习图像先验模型和正则化函数研究

Image restoration is a typical ill-posed problem, and regularization-based method is one of the commonly used methods. Although the image restoration algorithms based on the regularization method achieve state-of-the-art image restoration performance, these methods provide restored image with artifacts and less textures. In order to solve the above problems, this project carries out research from the following three aspects: 1) Constructing an image prior model combining semantic information, introducing high-level semantic information into low-level vision task, effectively expanding the spatial extent of the traditional image prior model, to improve the texture representation and ability of image prior model; 2) Designing a new regularization function, improving the expression ability of the function from the processing object and processing method, to improve the distinguishing ability of texture and noise; 3) Designing the training loss function, combining the image application, to get a more realistic image restoration model. The theoretical analysis and implementation of image prior model, regularization function and training loss function can promote the theoretical development of regularization based image restoration methods, and further broaden applications of image restoration in practice. Some of the previous research topics published in the top conferences of multimedia and computer vision and SCI search journals laid a solid foundation for the research work of this project.

图像复原是典型的不适定问题，正则化方法是常用方法之一。虽然基于正则化方法的图像复原算法取得了较高质量的恢复图像，但恢复图像中依然存在伪影以及纹理细节缺失的问题。为了解决上述问题，本项目从以下三个方面开展研究：1）构造结合语义的图像先验模型，将高层视觉中的语义信息引入到底层视觉任务中，有效扩大传统图像先验模型表示的空间范围，提高纹理表示和建模能力；2）设计新的正则化函数，从函数处理对象和处理方式上提升函数的表达能力，提升纹理和噪声的区分能力；3）结合图像分析应用，设计训练损失函数，得到更符合实际应用的图像复原模型。本项目关于图像先验模型、正则化函数、训练损失函数的构造和设计、算法实现和理论分析，可推动图像复原中正则化方法的理论发展，有助于拓展图像复原在实际中应用。部分前期已发表于多媒体和计算机视觉领域的顶级会议和SCI检索期刊上的研究内容为本项目的顺利开展奠定了坚实基础。

本项目面向图像复原问题，研究内容主要从以下三方面开展：1）结合语义信息的图像先验模型，构造新的图像先验模型，并尝试给出了一种近似求解方式。通过实验验证了结合图像分割的图像先验模型（S2IPM）可提升恢复图像视觉质量；2）结合不动点迭代和视觉Transformer模型，设计隐式正则化函数。通过实验验证了隐式通用正则化函数的有效性，一个预训练模型可处理4类图像复原问题的171个任务，同时能够取得与对比方法相近或者有竞争力的恢复性能；针对特定图像复原问题微调后，微调模型可到达与针对该问题训练的模型相同的恢复性能，有效提升了基于视觉Transformer的图像复原模型的训练效率；3）结合应用场景，设计图像复原监督训练的损失函数。通过实验验证了结合应用的损失函数训练方法（CASP）可提升恢复图像视觉质量。同时，恢复图像用于联合训练的物体检测模型时可提升检测模型对于噪声的抗干扰能力。