Aa wavelet model of ganglion cells array and its application in image representation

In this paper, we explore a new local image descriptor based on the modeling of ganglion cells array at the retina. We first introduce the mathematical model of a single ganglion cell and detailed distribution characteristics of ganglion cells array. From above evidence, we find out that the features of ganglion cells array are similar with wavelets in nature. Hence, a set of novel wavelet basis functions is constructed, which well fit these features. Furthermore, we discover that the modeling wavelets are in possession of tight frame property in its corresponding spatial domain expression, that is, multiple scales, shifts, phases of the spatial wavelet basis functions form a tight frame of the L2 Hilbert space, which is helpful to extract strongly distinctive and slightly redundant information in the image matching task. Finally, we obtain an efficient descriptor, state-of-the-art ones. Additionally, the proposed descriptor is easier to construct and much faster to compute. Evaluated in the image matching task on the Multi-view Stereo Correspondence Data set, the results demonstrate its effectiveness.