Towards Image Retrieval for Eight Percent of Color-Blind Men Vassili A.Kovalev Centre for Vision, Speech and Signal Processing, University of Surrey, Guildford, GU2 7XH, UK United Institute of Informatics Problems, Surganova 6, 220012 Minsk, Belarus v.kovalev@surrey.ac.uk Abstract About 8% of men (but not women) are suffering from color blindness. The objective of this work was to inves- tigate the problem of image retrieval based on color co- occurrence features when comparing normal vision and three kinds of color blindness (dichromasia): protanopia, deuteranopia, and tritanopia. Original database comprises 12000 images that were also converted into three dichro- matic versions using the Vischeck simulation tool. Results of 48000 queries were used to study influence of color blind- ness on retrieval results. Principal Component Analysis, Multidimensional Scaling, Hierarchical Clustering, Sup- port Vector Machines, and statistical methods were em- ployed for investigating feature space distortions associated with color blindness. Keywords: image retrieval, color blindness, color co- occurrence, correlogram 1 Introduction Color has been an active area of research in image re- trieval, more than in other branches of computer vision, due to the superior discriminating potentiality compared to gray-levels [16]. There is a great body of literature on color image retrieval including such particulars as color quantiza- tion, influence of illumination, color constancy, color fea- tures, and improvements over the traditional RGB space coming with opponent color representations (e.g., a com- parative study of two kinds of features and eleven color spaces [15]). However, there are several factors of biomed- ical and social nature that may have even stronger impact on this vibrant topic of research and developments. For in- stance, there are certain differences in color-perception be- havior associated with gender: women have more rich color vocabulary [14], they are better in matching color from memory [12], and they also showed greater responsiveness to the long-wavelength segment of the spectrum in color discrimination tasks [11]. More importantly, about 8% of men (but not women) are suffering from color deficiency which is also known as color blindness or dichromasia [19], [13], [2]. Color blindness is the inability to distinguish dif- ferences between certain colors. The most common types of blindness are protanopia (the spectrum is seen in tones of yellow and blue) and deuteranopia (confusion of red and green). Relatively rare is the tritanopia (spectrum is seen in tones of red and green). Contrary to such pathologies as achromatopsia, which comes from head injury or stroke and erases all experience of color, the color blindness is an incurable, genetic condition. The perception of color in dichromats is substantially different relative to people with normal vision. Thus, the objective of this work was to in- vestigate the problem of image retrieval based on color fea- tures alone when comparing normal vision and three kinds of color blindness: protanopia, deuteranopia, and tritanopia. Similar to other areas (e.g., [18],[13]), this is the first step towards making the life easier for dichromats in the digital world. 2 Image data Original image data comprises twelve thousand good quality color RGB images (computer wallpaper pho- tographs) of wide semantic diversity. By convention they are subdivided into 41 categories such as animals, art, avia- tion, birds, cars, history, food, fantasy, gifts, insects, money, machines, mountains, people, sea, patterns, trains, etc. The original image size of 1024 768 pixels has been reduced by a factor of two to 512 384 for convenience. Each of three types of color blindness was modeled by converting the ”normal” database into its dichromatic versions using the commonly accepted Vischeck dichromat simulation tool that implements algorithms developed in [2]. Hence, the total number of images involved in the present study was 12000 4=48000 (27 GB, uncompressed). 0-7695-2128-2/04 $20.00 (C) 2004 IEEE