Sub-regional fuzzy features for image retrieval

The use of pre-attentive sub-regions of a face instead of the whole face while senrching 'similar' faces in an image database is investigated. This is especially relevant in situations where the face is artificially distorted using make-up or disguise to avoid detection. It has been found that the sub-region facial features are much more effective in ordering similarity of faces compared to that obtained with whole face features in the sense that if one feature in a sub-region fails for some reason, other features may fulfil the task. Individual ranking of each sub-regional features is utilized as truth values of fuzzy predicates in similarity measure and the Combined Rank Set produced very good result in image retrieval process. Use of fuzzy predicate emphasizes the interrelations between the regional features in a face and thus re-occurrence of an element in the combined rank set can be explained as siren!: recommendation of that feature in similar faces. The storage space may be reduced in a large image database by storing only pre attentive features of sub-regions in vector form instead of storing all sub-regional features in matrix form. Experimentally, it is shown that these pre-attentive features call be used for image query in a personal computer with a huge saving In storage space - as large as 128:3 for face images of size 64 by 64.