基于稀疏表达的森林火焰及烟雾多层次奇异性特征识别机制研究

Due to the obscurity boundary of features of flame and smog in a forest image captured by a moving tripod head in complex natural environment, the region rapidly orientated and accurately recognized are difficult in currently used color spaces and available classification methods. The feature space's orithognal transfermation making them with less ralativity is reasonable and consistent with image sparse representaion. This project aims at revealing the multi-level sparse representation and classification mechanism of flame and smog, probes image recognition algorithm based on multi-level sparse representation, and finally provides proposals methodolog for the design of forest fire detecting system in complicated conditions. The project is going to be undertaken in 3 aspects as followings:.(1) Do research into the color transformation and the textures correlation of flame and smog, and compare the properties between each channel in available color spaces. The definition and transformation of each channel of flame and smog and the selection and fusion of their features depend on the method used for processing data in higher-dimension space in theory. The orthogonalization of the color and feature space makes the boundary of each variable demarcated through using principal component analysis with minimum correlation,maximizing the entire divergence between the eigenvectors and minimizing the correlation between the eigenvectors. Using feature selection algorithm of generalization error based on localized regularization makes the selection of feature more reasonable, and using nonnegative matrix decomposition based on linear projection structure gains clearer relationship between features. By means of color transferring ,the colors have a smaller correlation with orthogonality..(2) Do research into image recognition methods based on multi-level sparse representation of color, edge and texture, use discrete stationary wavelet, curvelet and wave atoms to research feature description and sparse representation of local and global invariance of color vectors. Specially do research into the rapid construction methods ofthe sparse dictionary by using K-SVD and probe genetic algorithm and block calculation,research rapid classification mechanism and comparison methods of multi-level sparse representation of the integral structure and feature of flame and smog..(3) Do research into the comparison of methods of rapidly orientating and classifying region of flame and smog, research into the construction and orientation mechanism based on feature hierarchical and decomposition of flame and smog sample.The initially orientation of flame and smog region are implemtented by PSO and GA, and then the exactly orientation, classification, comparsion and verification are finished by sparse representation. And compare this method with the used developing methods such as SVM, BP neural network,AdaBoost and so on.

针对复杂环境下运动云台摄制的森林图片中的火焰及烟雾的各特征数据的界限模糊，使得在现有彩色空间和分类方法下仍不能精确快速地对其进行定位和识别现象，引起人们对其特征空间定义、变换、融合和用多层次稀疏整体表达和分类机制产生研究兴趣。本研究围绕火焰及烟雾在新的彩色空间的自适应正交变换方法展开，探索基于颜色、边缘和纹理等特征的多层稀疏表示的图像识别算法，进行火焰及烟雾区域定位和识别方法的比较性研究，为复杂环境下森林火灾识别系统的设计提供方法和指导。重点研究：（1）对火焰及烟雾的在彩色空间的自适应变换方法和颜色奇异性特征描述的研究；（2）探索火焰及烟雾的多层次稀疏表达和分类机制和方法；（3）探索基于颜色、边缘和纹理等的多层稀疏表示的图像识别算法，（4）完成与稀疏表达相关的图像特征和分类的各种方法比较研究；（5）建立火焰及烟雾区域定位和识别方法的评价体系，为新型森林火灾快速识别系统的设计和集成奠定基础。

复杂环境下室内和室外视频图像中的火焰及烟雾的各特征数据的界限模糊，使得采用现有彩色空间和分类方法下仍不能精确快速地对火灾区域进行定位和识别，现有的特征空间定义、变换、融合和用多层次稀疏整体表达和分类机制的方法仍然需要进行更深入的研究。本研究项目已围绕火焰及烟雾在新的特征空间的变换方法展开了积极有效的研究，探索了基于颜色、边缘和纹理等特征的多层稀疏表示的图像识别算法，进行了火焰及烟雾区域定位和识别方法的比较性研究，为复杂环境下火灾识别系统的设计提供方法和指导。该研究重点研究了：（1）对火焰及烟雾的在不同彩色空间的自适应变换和奇异性特征描述方法进行了研究，提出在CMYK颜色空间对火灾区域用粒子群优化快速定位方法，定位的效率和精确度大大提高。用Relief特征选择、PCA和KPCA使提取和组合的特征更加具备特异性和可区分性，提出用不同时序下的方向梯度直方图HOG动态特征表达火焰和烟雾的运动属性, 提出用HSV 彩色模型形象化地表达光流分布。（2）提出用协方差算子描述火焰和烟雾的各特征间的相互关系和组合方法，用样本库和测试样本间的协方差算子的黎曼距离的变化判断特征的重要性，探索了这些奇异特征描述的火焰及烟雾的多层次稀疏表达和分类机制和方法；（3）探索基于颜色、边缘和纹理等的多层稀疏表示的图像识别算法，主要研究了光流直方图HOF和方向梯度直方图HOG的特征词袋字典表达方法和稀疏表达分类的方法。（4）完成与稀疏表达相关的图像特征和分类与各种诸如SVM, BP神经网络,AdaBoost和随机决策森林方法的比较研究，主要研究了KSVD,MP,OMP的稀疏字典构造和稀疏表达分类方法并编写高级语言进行了编码效率比较。（5）为新型火灾快速识别系统的设计和集成奠定了基础，研究和设计了针对不同特征变换、定义和包括稀疏表达分类的各种方法构成的系统软件框架。