Statistical modeling of the Gabor filter magnitude using Gamma distribution for effectively vehicle verification

Vehicle verification based on still image feature can be considered as supervised classification problem. An image descriptor is directly derived from the Gabor filtered output statistics of a given image. In general, the magnitude of the Gabor filtered output is modeled as the Gaussian distribution. So that the image descriptor is composed from mean, standard deviation, and skewness value of the Gabor filter magnitude [5, 6, 8]. However, Arrospide et. al. [9] argued that the skewness parameter is not meaningful for the class separation. Then, the feature descriptor is well defined only using mean and standard deviation of Gabor output distribution which leads to lower feature dimensionality. Based on our observation, the magnitude of the Gabor filter has strong tendency to follow the Gamma distribution. We propose a new texture descriptor derived from the maximum likelihood estimation of the Gamma distribution for effectively vehicle verification task. Experimental result shows that the proposed method is superior to the former approach under several classifier techniques.