Texture classification using improved 2d local discriminant baisis

Abstract

Local discriminant basis (LDB) is a supervised feature extraction method. Since LDB uses a redundant dictionary of wavelet packet basis, it is an excellent feature extractor for underlying features on time-frequency plane. Extraction of relevant features is a very important issue in signal and image classification. This reduces the computation complexity , while increasing the classification accuracy in the absence of adequate training samples. LDB algorithm utilizes best basis algorithm to find most discrimination basis among orthonormal redundant basis provided by wavelet packet transform. The classification performance of LDB algorithm is further improved using optimally weighted features. The genetic algorithm is utilized to find the optimal weight coefficients. The optimization step decreases within-class variance of training textures and increases between-class distances of teacher textures. The proposed algorithm is then tested on problem of composite texture segmentation. The results are compared with the weight coefficients obtained from Euclidean distance, J-divergence and Hellinger distance. Misclassification error is reduced by 26%-31%.