Comparison of color representations for content-based image retrieval in dermatology

Abstract

We compare the effectiveness of 10 different color representations in a content-based image retrieval task for dermatology. As features, we use the average colors of healthy and lesion skin in an image. The extracted features are used to retrieve similar images from a database using a k-nearest-neighbor search and Euclidean distance. The images in the database are divided into four different color categories. We measure the effectiveness of retrieval by the average percentage of retrieved images that belong to the same category as a query image. We found that the difference of the colors of lesion and healthy skin is a better color descriptor than the pair of these colors. We obtained the best results with the CIE-Lab color representation [75+/-3.8% (95% confidence interval) correct retrieval rate for k=11], followed by CIE-Luv and CIE-Lch. CIE-Lab is the most effective color space for content-based image retrieval of dermatological images. The difference of the colors of lesion and healthy skin in an image is a better color descriptor than the pair of these colors.