快速单像素多目独立成分设计

分析了LCNN的约束项的物理意义，认为约束项λ是有监督学习的加速度，使得整个算法无论是学习矩阵还是独立成分的求解效率都可达到O(n)。针对不同的λ和源信号、观测信号对的不同特性，提出了4种快速LCNN算法，分析了静态图像独立成分分析模型，建立了单像素内的独立模型，并总结了其优势。
关 键 词 Helmholtz自由能; 独立成分分析(ICA); Lagrange约束神经网络; 单像素内独立成分模型

Abstract The traditional independent component analysis is based on statistics mean of all aposteriori data and dismissed geometry. The classical lagrange constraint neural network (LCNN) employs Helmholtz free energy to unify supervised and unsupervised learning, and uses aprior and multi-sensing to solve independent components in one pixel, whose geometrical grain reached single pixel. The operations among pixels can be completely run parallelly. However, the constraints of classical LCNN bring ill-conditional matrix. In this paper, the real meaning of the constraints λ is discussed, four fast LCNN algorithms are proposed, the independent component (IC) models of still image is analyzed, and a new sub-pixel IC model is presented.
Key words Helmholtz free energy; independent component; lagrange constraint neural network; sub-pixel independent component model