Feature determination and inexact matching of images of industrial components

Techniques for the segmentation and recognition of images of overlapping industrial components in visually cluttered environments are described. Image segmentation is based on the Hough transformation of preprocessed edge data, employed as a predictive mechanism for line fitting, merging and redundancy checking, coupled with feedback verification. In the recognition phase, the extracted features are matched inexactly against a database of known models. Hypotheses are based on local feature matches and verified, first, by comparison of a translated and rotated prototype with the feature data, and, secondly, by re-examination of the original gradient data. Examples of experiments performed on single and multiple-component images are presented, demonstrating the robustness of the method.