Spatial local binary patterns for scene image classification

Local binary patterns (LBP), which is an effective and efficient texture descriptor, has been successfully applied to image analysis tasks such as face recognition and scene categorization. However, the conventional LBP histogram ignores spatial information of objects existing in images. In this paper, to address this problem, we propose a new Spatial Local Binary Patterns (SLBP) approach to encode geometric information of objects within images. The proposed method focuses on two critical aspects of scene classification: SLBP descriptor for image representation, and kernelized classifier for categorization. For the former one, SLBP descriptor is used to generate a series of ordered LBP histograms for capturing spatial information, by projecting LBP descriptor of an image onto different resolution and directions by linear projection, or points by circular projection. For the latter one, in order to fuse information from various resolution and direction, the support vector machines classifier with linear combinations of kernels is learned in different resolution and direction for scene image classification. Experimental results conducted on two popular benchmark datasets show that the gain over the original LBP may attain up to ten percentage points, which demonstrates the effectiveness of the proposed SLBP technique.