Decorrelation of a Sequence of Color Images Through Hierarchical Adaptive Color KLT

In this work is presented new algorithm for Hierarchical Adaptive Color Karhunen–Loève transform (HACKLT) aimed at the decorrelation of sequences of color RGB images. To increase the decorrelation, on each triad of components of same color, obtained from three consecutive images, is applied Color KLT (CKLT). To achieve maximum decorrelation, for each three sequential color images is detected the optimum orientation of the 3-component vectors used for the calculation of their covariance matrix of size 3 × 3, in one of the directions x, y or z. The so chosen orientation enhances the diagonalization of the covariance matrix. The hierarchical organization of the algorithm permits the calculations to stop in the level, in which the covariance matrix is diagonalized. Besides, by using HACKLT is achieved very high power concentration in the first eigen images of the decomposition. Compared to the “classic” KLT, used for the vectors which represent a sequence of images and are oriented along the direction “z” (for example, time), the computational complexity of the new algorithm is twice lower. The high efficiency of HACKLT in respect of the decorrelation defines its most suitable application areas: video compression, computer vision, extraction of features for objects recognition, objects tracking in video sequences, etc.