2D invariant object recognition using Log-Polar transform

Object recognition is an essential task in many image processing applications. Although object appears as three-dimensional in real world, they are usually perceived as two-dimensional in digital image or video. In most cases, major problems in recognizing objects lie on the two-dimensional geometry changes in object appearances. This paper presents an innovative template matching based object recognition method that is invariant to rotation and scale changes as well as resistant to noise. The approach is achieved by combining feature based search strategy and object matching in Log-Polar domain. Translation of the object is recovered with the new Gabor feature extraction method applied in the Cartesian coordinate. The multi-resolution Log-Polar search method is invented to reduce the number of feature point for Log-Polar matching in the target image. The new similarity measure for classification and verification is also proposed. The innovative combination of these techniques yields robustness in object recognition for fast computation without any rescaling in the target image. Comparisons of the repeatability factor of the new Gabor feature extraction with other well-known techniques such as SIFT and Harris-Laplacian is presented to evaluate the two-dimensional invariant and noise resistant properties of the proposed feature extraction. Experiments with still images and noisy images are provided to verify the effectiveness of the proposed approach in practice.