Refined Gaussian Weighted Histogram Intersection and Its Application in Number Plate Categorization

This paper proposes a refined Gaussian weighted histogram intersection for content-based image matching and applies the method for number plate categorization. Number plate images are classified into two groups based on their colour similarities with the model image of each group. The similarities of images are measured by the matching rates between their colour histograms. Histogram intersection (HI) is used to calculate the matching rates of histograms. Since the conventional histogram intersection algorithm is strictly based on the matching between bins of identical colours, the final matching rate could easily be affected by colour variation caused by various environment changes. In our recent paper [9], a Gaussian weighted histogram intersection (GWHI) algorithm has been proposed to facilitate the histogram matching via taking into account matching of both identical colours and similar colours. The weight is determined by the distance between two colours. When applied to number plate categorization, the GWHI algorithm demonstrates to be more robust to colour variations and produces a classification with much lower intra-class distance and much higher interclass distance than previous HI algorithms. However, the processing speed of this GWHI method is still not satisfying. In this paper, the GWHI method is further refined, where a colour quantization method is utilized to reduce the number of colours without introducing apparent perceptual colour distortion. New experimental results demonstrate that using the refined GWHI method, image categorization can be done more efficiently.