A robust texture feature extraction using the localized angular phase

Abstract

This paper proposes a novel descriptor, referred to as the localized angular phase (LAP), which is robust to illumination, scaling, and image blurring. LAP utilizes the phase information from the Fourier transform of the pixels in localized polar space with a fixed radius. The application examples of LAP are presented in terms of content-based image retrieval, classification, and feature extraction of real-world degraded images and computer-aided diagnosis using medical images. The experimental results show that the classification performance of LAP in terms of the latter application examples are better than those of local phase quantization (LPQ), local binary patterns (LBP), and local Fourier histogram (LFH). Specially, the capability of LAP to analyze degraded images and classify abnormal regions in medical images are superior to those of other methods since the best overall classification accuracy of LAP, LPQ, LBP, and LFH using degraded textures are 91.26, 61.23, 35.79, and 33.47%, respectively, while the sensitivity of LAP, LBP, and spatial gray level dependent method (SGLDM) in classifying abnormal lung regions in CT images are 100, 95.5, and 93.75%, respectively.