A novel translation, scale and rotation invariant feature extractor and its applications to target recognition

Feature extraction constitutes the basis of pattern classification and recognition. Based on the property of local entropy of describing local image properties, a multi-window binary local entropy based feature extraction algorithm is proposed. By normalizing the target using its geometrical moment, the feature vectors have translation and scale invariance, and the circle local window is introduced to make the feature vectors be rotation invariant. Feature extraction and recognition are performed with 12 airplanes in standard pattern library and real-world infrared target. The simple calculation, insensitivity to noise and superiority of multiwindow binary local entropy over moment invariants and Zernike moments are experimentally verified.