A Novel Feature Space for Classifying Textures and Objects in Low-Resolution Infrared Images

Abstract

This paper reports formation of a novel feature set for classifying textures in low-resolution thermal infrared (TIR) images like the ones acquired during aerial and ground operations of robotic vehicles. The proposed 3-component feature set includes energy coefficients obtained via 3-level overcomplete wavelet decomposition of subimages; three compact statistical descriptors derived from the grey-level co-occurrence matrices of TIR images and; a fractional energy descriptor ?. The energy descriptor ? accounts for emissivity related grey-level variations in the imaged object’s surface. Thus ? would provide succinct information about the influence of the imaged surface characteristics (shape, ambience and tidiness) on grey-level distribution in the image/surface. A fuzzy K-nearest neighbor classifier was used for labelling the image vectors. The reported results show that the proposed feature space would be helpful in classifying textures acquired from a distance under difficult illumination conditions.