图像块结构提取与分析及其在图像分类中的应用

Fast generations of massive image data are pushing the development of intelligent image understanding technology, and the extraction and analysis of image patch structure should be its basis. Based on the methods including Total Variation, Laplacian Pyramid, Multiple Resolution, W Orthogonal Function Decomposition and Image Depth Estimation, we would carry out the research on three aspects: (1) extract and analyze the information of image patch structure; (2) statistically analyze patch structure and details; (3) construct two models, Bag-of-Words model and a model of Convolution Neural Networks associated with a Bayes net, based on the local and global information formed by statistical information of patch structure and details. Realize automate labeling of image patch and detect the fire case of a forest scene image or video based on the supervised machine learning method. The contributions are :. (1) gradient image sparse representation based on an orthogonal function system, W system;. (2) extraction of image patch structure based on the combined information of depth and edge details;. (3) calculation of statistical information of details for a large scale image based on image patch structure;. (4) patch-based image labeling for Bag-of-Words model and deep learning associated with a Bayes network..The distinguishing features of the proposal are: balance of the local patch information and the global image information; integration of image pixel color, texture and geometry structure; combination of the basic theory and application on forestry industry.

海量图像数据的快速产生推动着图像智能理解技术的发展，而图像块(Patch)结构提取与分析是图像智能理解的基础。以约束全变差理论、拉普拉斯金字塔技术、多分辨率技术、W正交函数系分解和深度计算成果为基础，本项目主要研究：(1)图像块结构信息的提取与分析；(2)块结构细节信息的统计分析；(3)依据块结构和细节统计信息形成的局部与全局信息，构造词包模型和级联Bayes网络的卷积神经网络模型，基于有监督的学习实现图像的自动标注和森林烟火场景的识别。创新点体现在：基于 W正交函数系的梯度图像稀疏表示；融合深度信息和边缘线细节信息的图像块结构提取；基于块结构的大尺度图像细节统计信息计算；词包模型、深度学习级联Bayes网络的图像按块标注。项目研究主要特色在于：局部信息与全局信息相统一；颜色纹理信息和几何结构信息相统一；基本理论研究与林业行业应用相结合，实现把信息理论新技术应用于林业管理。

随着手机拍照和摄像头的广泛使用，图像数据呈爆炸式增长。为了推进对图像数据的快速且智能的理解，本项目在图像融合、图像增强和图像去雾等研究的基础上，开展了图像块(Patch)结构提取与分析方面的理论研究，并把理论研究成果应用于图像分类应用中。项目研究取得的创新成果包括：提出了以保边和提高计算效率为目标的多个双向滤波器，并以此为基础提出了图像块结构提取的多个方法；提出了在图像细节统计分析的基础上，把W正交函数系与经验模态分解、非子采样方向滤波器、暗原色先验等多种方法融合的图像融合、图像增强和去雾算法；构建了非负矩阵分解的优化问题和求解方法，提出了基于深度学习、矩阵非负分解为基础的图像数据和点云数据的分析和分类方法。上述研究成果，对于图像理解分析、图像数据管理和基于监控的视频分析都能够提供有效的技术支持。