IMAGE SEGMENTATION WITH LOCAL FEEDBACK AT LOW- AND HIGH-LEVEL PROCESSING (COMPUTER VISION, CONTOUR EXTRACTION, REGION GROWING, HANDWRITTEN NUMERAL RECOGNITION).

摘要

The purpose of this work is to explore approaches other than using top-down feedback from the interpretation subsystem to the segmentation subsystem, for improving the segmentation results and thus the reliability and performance of a general computer vision system.;We presented two effective general-purpose low-level image segmentation algorithms: the continuity-preserving contour extraction algorithm and the boundary-bounded region growing algorithm. We studied the use of local feedback loops in the segmentation subsystem as well as in the interpretation subsystem. We suggested that a more general and reliable architecture for both segmentation and interpretation subsystems should consist of not only a basic hierarchical algorithm but also a local feedback loop across it. Through the local feedback loop in the segmentation subsystem, the segmentation results can be monitored at any time and refined whenever necessary before they are sent to the high-level stage of interpretation. On the other hand, through the local feedback loop in the interpretation subsystem, many segmentation errors can be corrected locally in the interpretation stage without having to feedback to the low-level segmentation stage. We applied such architectures to a realistic application: a handwritten numeric character recognition system, which can recognize not only isolated numerals but also touching and broken numerals. As expected, this was made possible by the incorporation of local feedback loops in the segmentation and recognition subsystems. As a conclusion, we discussed a plausible architecture for the general computer vision system.