Diatom recognition by convex and concave contour curvature

We describe a new contour feature set. A contour is segmented into convex, concave and straight segments, after which length and curvature features are computed. A symmetry analysis allows the detection of the number of elementary segments. Results on two contour test sets were obtained: Using only four features, a simple nearest-mean classifier yielded a perfect identification (ID) rate of 100% on a small set consisting of shapes with minute differences, which are difficult to identify even for human experts. Using 10 features, it yielded 83.5% on a large set with very diverse shapes.