An Efficient Sparse Data Filtering Method for Image Histogram Comparison

The traditional method for creating a color histogram is to equally subdivide a certain color space into a given number of bins and then count the number of pixels each bin contains. Using such a histogram, the performance of almost all methods of histogram comparison are highly sensitive to the localization of the bin boundaries. In order to alleviate this sensitivity a new method of histogram creation is proposed in this paper. The histogram at a certain resolution (number of bins) is obtained in two steps. First a low pass filter is applied onto the highest resolution histogram and then the filtered histogram is quantized into the required number of bins. Based on the sparse nature of the color histogram an efficient algorithm which concomitantly performs both steps is proposed. The algorithm does not require the prior creation of the highest resolution histogram. It creates the required histogram directly in a single scan of the original image by counting the contributions of each image pixel to different histogram bins. The experimental results reveal a significant improvement of image retrieval performances when the histograms are created using the proposed method over the case they are created using the traditional approach.