数据驱动图像处理中基于图像属性的样例集选择与学习

Images / videos with massive, diverse and unstructured features brought great challenges to classic image processing methods developed since 1950’s. Data-driven image processing, rising from the beginning of this century, greatly enhanced the performance of the classical image processing methods. Data-driven image processing leverages advanced machine learning (ML) methods in order to learn the priori model from a set of collected image examples pertaining to image attributes. However, the performance and flexibility of data-driven image processing methods heavily rely on the choice of training examples, and there is not yet a systematic study of automatic selection of vast image examples for the training set. In the field of image processing, the training examples are chosen in a heuristic way for a specific task, or specific object. Meanwhile, machine learning techniques, targeting at classification tasks, select features or examples based on the correlations of class labels for non-image samples, significantly differing from images with higher dimensionality and versatile tasks. In this project, we address the issue of learning to select training examples based on image attributes ranging from the straightforward appearance and geometry, to intrinsic ones including sparsity and low rank. We intend to unify the learning of training sets for image enhancement, alignment, and other tasks by a general framework derived from partial difference equations. This study is able to yield processing algorithms of high performance adapting to various tasks and data, and to enrich the theory and practice for example selection in machine learning.

具有海量、多样和非结构化特点的图像/视频数据大量涌现，给上世纪50年代以来的经典图像处理手段带来极大挑战。本世纪初兴起的数据驱动图像处理，借助机器学习从样例集学习反映图像属性的先验模型，极大提升了经典图像处理方法的性能。然而，数据驱动图像处理方法的性能和适应性严重依赖于训练样例的选择，目前尚未有从大量图像数据中自动学习训练样例集的系统研究：在图像处理领域，常规思路仅根据特定任务、特定对象人工选择训练集；机器学习领域针对分类任务，选择非图像样例集，其数据维度与任务复杂性与图像差异很大。本项目从特定的图像增强、图像对齐到通用图像处理，从直观的图像表观属性、几何属性到抽象的图像内在属性（如稀疏和低秩）等3个方面，深入研究训练样例集自动选择和学习方法。通过项目研究，探索建立适应多样任务的统一图像样例集选择与学习框架，以获得相应高性能处理算法，并在一定程度上丰富机器学习领域数据选择的相关理论和实践。

本世纪初兴起的数据驱动图像处理，借助机器学习从样例集学习反映图像属性的先验模型，极大提升了经典图像处理方法的性能。然而，数据驱动图像处理方法的性能和适应性严重依赖于训练样例的选择，目前尚未有从大量图像数据中自动学习训练样例集的系统研究：在图像处理领域，常规思路仅根据特定任务、特定对象人工选择训练集；机器学习领域针对分类任务，选择非图像样例集，其数据维度与任务复杂性与图像差异很大。本项目针对图像去雾、去模糊，三维重构，无监督深度估计，深度神经网络理论研究这四个方面，重点研究基于“知识”+“数据”的图像除雾、去模糊模型，基于不变特征基本理论的三维重构模型，基于深度估计的无监督深度学习架构模型，以及具有数学原理支持和保证的深度神经网络模型。