USING COLOR HISTOGRAMS FOR OBJECTS DETECTION IN RESIZED AND ROTATED IMAGES

Abstract

The objects detection in images is based on a set of their features, the nature of which can be very different. In general, these features are divided into three levels: the bottom, which includes color and texture; middle ‑ contours/outlines and upper ‑ semantic and statistical characteristics. There are methods based on the analysis of intensity histograms of individual objects, both in monochrome and color versions. The most commonly used color histogram in digital image processing is RGB. However, those being closest to the spectral representation of color are HSL, HSV, and HSI (general notation HS*), which describe the color in a cylindrical coordinate system, where the color is an angular coordinate.

This work analyzes the possibility of using the color histograms in the *.bmp and *.jpg formats generated in HS* color spaces for content based retrieval. Histograms analysis was carried out by their comparison via the cross-correlation coefficient R calculation. The studies were performed for Caltech 101 dataset standard samples and also for a set of own test images.

In the objects recognition tasks, the size of the objects in the image is often different from the size of the template. Therefore, we attempted to evaluate the limits of using H-histograms for comparing images with wide variations of their sizes. Studies of the change of the cross-correlation coefficient R when the image size was reduced by 2ⁿ times (n = 1…9) show that the use of the H-histogram allows to successfully recognize the image when its size is reduced up to 128 times.

Also, the dependence of the H-histograms cross-correlation coefficient R for the test images during their rotation was studied. The rotation angle was varied from 0^o up to 180^o. Slight deviation of the cross-correlation coefficient R value from unity for most test samples allows to assume the possibility of using the proposed method not only to compare compressed images but also rotated ones.

The research results show a possibility of the H-histogram using for comparing the color images and, subsequently, recognizing their fragments due to factors that impair image quality, that allows this approach to be used in Content Based Image Retrieval (CBIR) systems.

In addition, the research results indicate the further possibility of constructing a color image descriptor based on the H-histogram.

Key words: color histograms, cross-correlation coefficient, HS* color system, content based image retrieval.