Translation, Scaling, Rotation Invariant Descriptor Analysis Approach To Object Recognition Based On Moment Invariants

Object recognition is a very important problem in computer vision and digital image processing. Invariant moment has found extensive application in the field of object recognition due to its ability of representing global features and characteristic independent of translation, scale and rotation. M. K. Hu first introduced seven invariant moments in 1961, and then invariant moment had widely used in object recognition, based on the method, researcher have proposed kinds of invariant moments such as complex moments and radial and angular moments. But those moments are highly sensitive to noise, and reconstruction is extremely difficult. Furthermore, the computational cost increases acutely with it's orders. Later, some researchers proposed orthogonal moments, such as Zeruike moments, Pseudo-Zernike moments, Legendre moments, Orthogonal Fourier-Mellin moments, Tchebichef moments and Krawtchouk moments. Those moments present low information redundancy, highly noise robust, and images can be easily reconstruct from a set of orthogonal moments. This paper addresses the issue of selecting translation, scaling, rotation invariant features for object recognition use Radon transform(RM) and Radon-Analytic Fourier-Mellin transform(RFM). And we also proposed a new orthogonal moments based on the Bessel function(BM) for image analysis and recognition. Theoretical and experimental results show the superiority of RM approach including low computational complexity and high robustness to additive noise, and better classification performance only using a very few descriptors; BM presents the same performance in object invariant recognition as the orthogonal moments-based methods, and low information redundancy and high noise robust.