Gray-scale Invariant Texture Classification Using Gradient Spectrum

A new gray-scale invariant feature extraction method for texture classification is proposed. In order to capture the local image texture along horizontal, vertical, diagonal, and anti-diagonal directions separately, we introduce four modified gradient operators called nonlinear gradient operators. The magnitudes of these nonlinear gradient operators are robust against any monotonic transformation of the gray scale. The joint distributions of these gradient magnitudes, which can be characterized using a histogram, can be used for texture classification and image retrieval. We call such a histogram approach gradient spectrum. The gradient spectrum technique only requires addition, division, absolute value, and counting operations, and thus is highly computationally efficient. The properties of gradient spectrum include translation and illuminant invariance, which is highly desirable in real-world applications. Gray-scale invariant texture classification experimental results show that the proposed method outperforms earlier gradient-based method.