Vector-based arc segmentation in the machine drawing understanding system environment

Arcs are important primitives in engineering drawings. Extracting these primitives during the lexical analysis phase is a prerequisite to syntactic and semantic understanding of engineering drawings within the machine drawing understanding system. Bars are detected by the orthogonal zig-zag vectorization algorithm. Some of the detected bars are linear approximations of arcs. As such, they provide the basis for arc segmentation. An arc is detected by finding a chain of bars and a triplet of points along the chain. The arc center is first approximated as the center of mass of the triangle formed by the intersection of the perpendicular bisectors of the chords these points define. The location of the center is refined by recursively finding more such triplets and converging to within no more than a few pixels from the actual arc center after two or three iterations. The high performance of the algorithm, demonstrated on a set of real engineering drawings, is due to the fact that it avoids both raster-to-vector and massive pixel-level operations, as well as any space transformations