Toward Generic Model-Based Object Recognition by Knowledge Acquisition and Machine Learning

摘要

The aim of this paper is to present a system that uses both inducted hypotheses and expert knowledge for recognition of objects within digital images. Classification of pixels by spectral and spatial features is learned via example. Recognition of objects within resulting class images is performed by a novel method of search that takes advice on object properties, and enables error in object recognition to be systematically isolated and rectified. Guided by the system, a user selects whether error correction is machine-learned or user-defined. The adaptability of mixed-initiative error correction leads to wide applicability. The system is demonstrated using images from a robot soccer domain. 1