Construct Abstraction for Automatic Information Abstraction from Digital Images

摘要

Automatic Machine Vision involves humans building machines capable of recognizing objects and scenes in digital images without further human assistance. Machine vision is a bottleneck in robotics and automated systems. When human programmers construct vision systems they are usually designed so that the program architecture and the data are optimized for the particular problem and classification technique being used. In general machine vision systems are hand-crafted to give the best results for a particular application, but are brittle and perform poorly outside their narrow specification, and lack any ability to adapt. On this project we have been researching a method of creating flexible machine vision systems that can modify their behavior and evolve in particular environments to recognize anything that an operator has indicated as being 'interesting' in that environment. For example, Figure 1 shows typical objects that a house-tidying robot might encounter during its everyday duties.