面向图像复原的高阶马尔可夫随机场先验模型研究

Due to the bad conditions in the process of imaging and transferring, the quality of image could be degraded, thus deteriorating its use. Image restoration is a technique of restoring and improving image quality, and it has been in focus all along. Image restoration is an ill-posed problem, however, through introducing some constraints, it can be turned into a well-posed problem. Markov Random Fields (MRF) model is a new method to model probability of prior constrains, and it has been used in image restoration. In order to satisfy the application needs of image restoration for multi-scene/multi-content/multi-degraded-model, this project focuses on image denoising and super-resolution, and researches new theory and methods of image restoration based on higher-order MRF prior model. This project tries to do the following researches: 1. Bring in higher-order non-spacial-adjacent neighboring structure, study MRF modeling adapted to context information in received image, construct new higher-order MRF prior model which merges multi-type structure features; 2. through analyzing the scene category/image content/degradation model of degraded image, build up sub-class analysis model of image restoration and corresponding training theory and methods; 3.through studying near-optimal acceleration solution and strategies for semi-blind model, provide high-practical image restoration methods based on higher-order MRF prior model.

由于图像成像条件和传输过程的不良因素，产生的图像质量下降影响了图像的使用。图像复原是恢复和改善图像质量的图像处理技术，受到人们普遍关注。图像复原是不适定问题，通过引入一定的约束条件可将不适定的图像复原问题转变为适定问题。MRF模型是一种对先验约束进行概率建模的新模型方法，开始应用于图像复原。本课题以图像去噪和超分辨率为研究对象，研究基于高阶MRF先验模型的图像复原的新理论和新方法，以高效适应多场景类别、多图像内容和多降质类型图像复原的应用需求。课题拟开展：（一）引入高阶非空间相邻的邻域结构，研究适应观测图像中上下文信息的MRF模型建模，构造新的融合多类型结构特征的高阶MRF先验模型；（二）通过分析降质图像的场景类别、图像内容与降质类型，建立图像复原的子类分析模型及其训练的理论和方法；（三）通过研究模型的快速次优解求解和半盲模型的解决策略，提供实用高效的基于高阶MRF先验模型的图像复原方法。

由于图像成像条件和传输过程的不良因素，产生的图像质量下降影响了图像的使用。图像复原是恢复和改善图像质量的图像处理技术，受到人们普遍关注。图像复原是不适定问题，通过引入一定的约束条件可将不适定的图像复原问题转变为适定问题。MRF 模型是一种对先验约束进行概率建模的新模型方法，开始应用于图像复原。本课题以图像去噪和超分辨率为研究对象，研究基于高阶MRF 先验模型的图像复原的新理论和新方法，以高效适应图像复原的应用需求。首先构建面向图像复原研究的样本图像库；其次分析了传统的专家场模型中所存在的不足，提出了一种具有边缘保持特性的专家场模型，构建了融合边缘特征进行图像复原的高阶MRF先验模型，获得了优于传统专家场模型的边缘保持效果；随后在传统专家场模型的迭代推断算法中，提出一种迭代求解的自动停止准则，提高了专家场图像去噪算法的去噪效果，并克服了传统专家场模型推断中必须假设噪声方差已知的前提要求，将专家场图像去噪方法较好的推广到实际噪声图像的去噪任务中；最后将高阶MRF先验引入图像超分辨率复原方法中，提出了一种基于专家场先验的图像超分辨率复原方法，扩展了传统基于低阶MRF的图像超分辨率复原方法的模型表述能力，改善了算法的超分辨率复原效果。