Gabor filtering is a well-known feature extraction method

which has been widely studied and applied in the field of machine vision. This paper presents a new multi-directional and multi-scale Gabor feature representation

extraction and its matching algorithm. By using a set of Gabor filters with different scales and different directions to filter an image

the filtered results in each direction are reorganized in the order of the scales and concatenated into a multi-directional and multi-scale Gabor feature. We further propose the concept of cyclic vectors and redefine a similarity measure for multi-directional and multi-scale Gabor features as the maximum similarity value between one feature vector and the corresponding cyclic vectors. Our experimental results show that the proposed descriptor not only has the characteristics of translational invariance

rotational invariance

and scale invariance

but also embody the good feature representation ability and the significant discriminative strength for the local region descriptors in image.