Continuous rotation invariant features for gradient-based texture classification

摘要: Four highly discriminative and continuous rotation invariant methods are proposed.We use the Hessian matrix Gaussian derivative filters.Verified on CUReT, KTH-TIPS, KTH-TIPS2-a, UIUC Brodatz texture datasets. Extracting features is a valuable technique for effective classification of texture. The Histograms Oriented Gradients (HOG) algorithm has been proved to be theoretically simple, applied in many areas. Also, co-occurrence HOG (CoHOG) provides unified description including both statistical differential properties patch. However, CoHOG have some shortcomings: they discard important information not rotation. In this paper, based original algorithms, four novel feature extraction proposed. first method uses filters named GDF-HOG. second third eigenvalues Eig(Hess)-HOG Eig(Hess)-CoHOG, respectively. fourth exploits means curvatures calculate image surface GM-CoHOG. We empirically shown that proposed extended provide useful invariance. results compared with algorithms KTH-TIPS2-a datasets show achieve best result all addition, we make comparison several well-known descriptors. experiments analysis carried out dataset, promising obtained from those experiments.