A Directional - Edge - Based Face Detection Algorithm Adapted to the VLSI Image Recognition System

There has been a considerable interest in the development of human-like image perception systems. Although a number of algorithms have been proposed so far, it is still a challenging task to achieve a human-like robustness in image perception under various circumstances. For robust image perception, the feature representation algorithm which extracts essential features from an image and represents them in a vector format is of critical importance. Since the discovery by Hubel and Wiesel (Hubel & Wiesel (1959)) in the study of visual cortex of animals, it is well known that the directional edge information in visual inputs plays an essential role in early visual processing. Namely, biological systems rely on the spatial relationship among edges in various directions for image perception. In this regard, local intensity gradients or directional edges have beenwidely used in computer vision and pattern recognition (Belongie et al. (2002); Dalai et al. (2005); Freeman & Roth (1995); Lowe (2004); Mikolajczyk & Schmid (2005); Shibata et al. (1999)). They are utilized in feature descriptors to represent object images in image and gesture recognition. The present work also relies on the edge information extracted from images. A feature vector representation algorithm called Projected Principal-Edge Distribution (PPED for short) (Yagi & Shibata (2003)) has been developed to meet the demand of neural-associative-processor-based hardware recognition systems (Shibata (2002)). The original idea of PPED was first disclosed in 1999 (Shibata et al. (1999)) and developed in the work of Yagi & Shibata (2003). The robust nature of the PPED vector representation has been successfully demonstrated through its applications to hand-written pattern recognition and medical radiograph analysis (Yagi et al. (2000); Yagi & Shibata (2002)). The feature representation in PPED is based on the spatial distribution of four directional edges (horizontal, +45◦, vertical, and −45◦) extracted from an image, in which the local variance of pixel intensities is taken into account for edge detection. Since the processing is computationally very expensive, a dedicated VLSI chip has been developed which can generate PPED vectors at a rate of 106 vectors/sec (Yamasaki & Shibata (2007)). An architecture that enhances the throughput by a factor of about 30 was also developed (Nakagawa et al. (2009)). VLSI chips specialized for the associative processing (maximum likelihood search) have also been developed in both digital (Ogawa et al. (2002)) and analog (Yamasaki & Shibata (2003)) technologies aiming at real-time performances. The face A Directional-Edge-Based Face Detection Algorithm Adapted to the VLSI Image Recognition System 11