Structured-light based joint recognition using bottom-up and top-down combined visual processing

In this paper a multilayer hierarchical visual processing architecture integrated with a bottom-up and top-down combined inference algorithm is proposed for robust weld joint recognition and localization. Three layers-pixel layer, primitive layer, and profile layer-are defined, firstly laser stripe centerline points are coarsely extracted from the image in pixel layer, then the primitive layer primitives are obtained by a grouping algorithm with a hypothesis-verification scheme, and at last a hypothesis for a joint pattern based on partial match is generated from profile layer and verified by searching through the lower layers of the hierarchy. During the top-down verification process, primitive that is partially extracted during former processing is recovered by using a local adaptive segmentation technique, which is intended to accommodate to different noise-to-signal levels. Experimental results validate the robust performance of this approach in the presence of heavy noise in real-time.