Understanding human vision processing has been the ultimate goal for many vision researchers, including machine vision developers. However, successes in specific problem domains such as face detection and gait recognition have driven the machine vision community to invest more effort on a so-called flat architecture than on a human-like deep architecture. The former employs a number of simple feature detectors compensated by a large amount of training data. The latter employs a hierarchy of specialized feature detectors to extract essential information from a smaller amount of training data. Through an extensive review of primate vision systems, the authors try to shift the trend toward the deep architecture.
展开▼
