Detection of Anomalies in Textures Based on Multi-Resolution Features

Multi-resolution decompositions, such as the wavelet transform, are often employed in anomaly detection algorithms for feature extraction. However, the extracted features may be unreliable for anomaly detection in textures due to inconsistencies between the assumed background model and the true data. In this paper, we present an anomaly detection scheme which relies on a statistical model of textures and is specifically designed for detection of anomalies in textures. Motivated by recent works on texture segmentation and texture classification, we introduce a multi-resolution feature space that facilitates anomaly detection with constant false alarm rate for a wide range of textures. Experimental results demonstrate that the proposed algorithm, when applied to images containing background texture, achieves improved detection results and lower false alarm rate than a competitive anomaly detection scheme.