A novel feature extraction method for image recognition based on similar discriminant function (SDF)

• Published in: Pattern recognition Vol. 26; no. 1; pp. 115 - 125
• Main Authors: Cheng, Yong-Qing, Liu, Ke, Yang, Jing-Yu
• Format: Journal Article
• Language: English
• Published: Oxford Elsevier Ltd 1993; Elsevier Science
• Subjects: Applied sciences; Artificial intelligence; Computer science; control theory; systems; Exact sciences and technology; Feature extraction; Hierarchical classifier; Human face recognition; Image recognition; Pattern recognition; Pattern recognition. Digital image processing. Computational geometry; Similar discriminant function; Hierarchical classifier; Feature extraction; Image recognition; Pattern recognition; Human face recognition; Similar discriminant function; Computer vision; Hierarchical classification; Information extraction; Algorithm; Artificial intelligence; Discriminant function
• ISSN: 0031-3203; 1873-5142
• DOI: 10.1016/0031-3203(93)90093-C

The extraction of image features is the most fundamental and important problem in image recognition. In this paper, a similarity measure of matrices is first presented, and a similar discriminant function (SDF) of images is established. Based on the discriminant function, we further propose a novel feature extraction method for image recognition. For each class of training image samples, an optimal projection axis maximizing the similarity among these training image samples for the class is calculated. Unlike the common methods of feature extraction, we extract a projective feature vector for a training image sample by projecting the image on the optimal projection axis of the class itself, and a set of projective feature vectors for a testing image sample by projecting the image on all the optimal projection axes. Finally, a hierarchical classifier in the optimal discriminant space is designed to recognize images. In order to test the efficiency of our method, it is used to recognize human faces and English characters. Experimental results have shown that our method has good recognition performance, and the extracted projective feature vectors contain more recognition information than commonly used image features.