Local Gradient Dual Coding Book (LGDCB) Framework for Effective Texture Classification

Abstract

Local descriptors are popular in texture classification and they are robust and sustain and perform well in varying pose, lighting and illumination condition. The accuracy of these local descriptors depends up on the precision of local features derived on the local neighborhood. To achieve robustness this paper initially computes eight directional edge responses using kirsch masks for each sampling point of 3x3 overlapping window. This paper divides the 3x3 edge responses( E r ) windows into a dual matrix consist of four sampling pixels each. On each matrix this paper computes the edge response ranks and based on this a coding book sequence number is generated for each matrix and this process derives dual edge response sequence matrix (DE r SM). The co-occurrence matrix and grey level cooccurrence matrix (GLCM) features derived on DE r SM represent the feature vector. The proposed descriptor is tested with well-known texture databases with different categories. The experimental results of the proposed descriptor are compared with state-of-art local descriptors and the results demonstrate the efficacy of the proposed method over the other methods.