A recursive segmentation and classification scheme for improving segmentation accuracy and detection rate in real-time machine vision applications

Abstract

The segmentation accuracy is shown to be a critical factor in detection rate improvement. With accurate segmentation, results are easier to interpret, and classification performance is better. Therefore, it is required to have a performance measure for segmentation evaluation. However, a number of restrictions limit using existing segmentation performance measures. A recursive segmentation and classification scheme is proposed to improve segmentation accuracy and classification performance in real-time machine vision applications. In this scheme, the confidence level of classification results is used as a new performance measure to evaluate the accuracy of segmentation algorithm. Segmentation is repeated until a classification with desired confidence level is achieved. This scheme can be implemented automatically. Experimental results show that it is efficient to improve segmentation accuracy and the overall detection performance, especially for real-time machine vision applications, where the scene is complicated and a single segmentation algorithm cannot produce satisfactory results.