50 Ioannis Anagnostopoulos et al.
ing the fuzzy logic rules segmented totally 841 skin areas. However, 30 faces were not selected and therefore the performance of this system is 93.77% (452/482). Following the fuzzy logic system, the ANN received the 841 skin areas and decided that 397 of them represent faces. Thus, the performance of the ANN is 87.83% (397/452). Finally, the overall system performance is 82.36%, since 397 from a total of 482 faces were identified. All the results are shown analytically in Table 5.
Precise Photo Retrieval on the Web 51
cluded. Nevertheless, the user got a significantly precise amount of information con-cerning the Boolean image query “Georgatos AND Inter”. Table 6 holds some results over a large sample from more than 300 mixed Boolean image queries collected by the three search services. Therefore, over 14857 images from which 8932 are indeed pho-tos, the FL rules successfully identified 8552 of them. It was evaluated that the failure in this step occurred due to the fact that the chrominance fuzzy logic rules do not cover sufficiently all the variation in terms of the human skin color. Other reasons come from the area of image processing involving limitations in illumination condi-tions or image texture and inconsistency. The ANN further processed the above set of images and decided that 8039 are actually photos that contain human faces, while 513 images were erroneously excluded from this class. On the other hand, this was ex-pected since many images depict human faces in different angle positions and not only frontal or slightly rotated. For similar reason and in accordance with the images that indeed present human faces, the procedure added erroneously 176 images, which are unrelated to the photo query submissions. However, using the ‘SearchPhoto’ meta-search procedure a significant large amount of irrelevant information was excluded as depicted in Table 6. Figure 7 shows the information reduction in case of irrelevant results as well as the information loss in case of the relevant results for the photos, the other retrieved images and the total returned meta-results.