Object Recognition under Changing Illumination

Let H be the histogram of the colors of the input image where H c) denotes the probability that color c occurs in the image. The centered moment is computed as follows  

Funt and Finlayson (1995) proposed histogramming of color ratios. They suggested computing the ratios between adjacent pixels. Let c = [cr, cg, ct]T and c = [c, c g, c h 1 be two colors that are obtained from two adjacent pixel positions. Assuming that the sensors are narrow band, the intensity measured by the sensor at position (x, y) is given by 

If we have applied the logarithm, i.e. p, = logc, then we can rewrite the Faplacian as 

Finlayson et al. (1995) report good results on a small database of 55 objects. The method seems to perform slightly worse than color constant color indexing that was developed by Funt and Finlayson (1995). 

Berwick and Lee (1998) suggested the use of the following chromaticity space for object recognition. They simply choose one of the color channels and divide the remaining channels by the chosen color channel. Let c = [cr,cg, C, T be the color of an image pixel. The chromaticity space is then defined as 

---