Face Recognition: Eigenface, Elastic. Matching, and Neural Nets. An introduction to the paper by Zhang, Yan, and Lades. People easily recognize one another ...
Prolog to
Face Recognition: Eigenface, Elastic Matching, and Neural Nets An introduction to the paper by Zhang, Yan, and Lades
People easily recognize one another by looking at each other’s faces. Recognizing another human’s face is such a fundamental task that even a toddler does it, and there is evidence that infants quickly learn to identify the faces of their parents. Hence, it is logical for us to imagine that a computer easily can be taught how to recognize an individual by looking at his face. As many researchers have discovered, however, such a “trivial” task is not simple for a computer to perform. After all, a computer learns like a computer and not yet like a human. Since computers are known for being capable of performing highly repetitive tasks, many efforts are still being made to “teach” them to recognize human faces. The automated human face-recognition tasks normally are formulated into two categories of problems: identification and verification based on human faces. Like other biometrics, the objective of an identification task is to determine the identity of an unknown human face, i.e., to answer the question, “Who is this?” In an identification task, a human face is presented and compared with each of the human faces stored in a data base, whose identities are known. Each comparison produces a similarity score, which indicates the degree of similarity between the pair of human faces compared. As a result, a matching candidate list can be produced in descending order of the similarity scores. In other words, the top candidate on the list is the identity associated with the stored face that produced the highest similarity score when compared with the presented face. This candidate list allows a human face-recognition expert manually to examine those faces that are most similar to the presented face to determine the identity associated with the presented human face. The objective of a verification task is to determine if the identity claimed to be associated with the human face really is the individual’s “true” identity, i.e., to find the answer to the question, “Are you whom you claimed to be?” In a verification task, a human face is presented together with the claimed identity for comparisons. The computer retrieves the stored human face (or faces) of the claimed identity and compares that with the presented face. Publisher Item Identifier S 0018-9219(97)06865-5.
A similarity score is again produced for each comparison. This similarity score is then used to determine whether the individual is of the identity he claims. An automated face identification or verification algorithm normally consists of three main components: 1) face finding (or locating), 2) face feature extraction (or representation), and 3) face feature matching. Zhang et al. focus their discussions on the last two components. The face feature extraction process normally produces a lower dimension representation of human faces. When selecting an appropriate representation (feature vector) for human faces, some important considerations are the discriminating power, the variance tolerance, and the data-reduction efficiency. The discriminating power is the degree of dissimilarity of the feature vectors representing a pair of different faces. The variance tolerance is the degree of similarity of the feature vectors representing different images of the same individual’s face. The data-reduction efficiency is the compactness of the representation. The face feature matching process compares the feature vector of a newly acquired image with that of one or more stored images with known identity for identification or verification purposes. The performance of three recently proposed twodimensional (2-D) (intensity) image-based approaches to face recognition (identification) are compared and analyzed in “Face Recognition: Eigenface, Elastic Matching, and Neural Nets,” and the results seem to indicate that elastic matching provides the best performance. Also discussed in this paper is an interesting three-dimensional (3-D) surfacebased approach proposed by Atick et al. This approach uses the face surface shape rather than the face image, thereby eliminating a serious problem plaguing almost all 2-D approaches: sensitivity to illumination. The 3-D face surface usually has to be obtained from the intensity image (shape-from-shading), however, which is a difficult problem. One may contemplate that in the near future, when a 3-D face surface can be obtained as readily as a 2-D face image (picture), a highly accurate face-recognition system could be developed based on such data. —Weicheng Shen and Rajiv Khanna
0018–9219/97$10.00 1997 IEEE 1422
PROCEEDINGS OF THE IEEE, VOL. 85, NO. 9, SEPTEMBER 1997
