Identifying Real-Time Hand-Sketched Engineering Drawing by Using RCR and NN

This paper introduced several new concepts such as sketch's centroid, centroid radii and regularized centroid radii (RCR), and proposed a hand-sketch neural network recognizer to distinguish 4 types of strokes, namely line segment, circle, half circle, and quarter circle. The approach constructed a hand-sketch classifier through extracting the sketch primitives' RCR as features, crossing the four primitives' RCR values as the learning samples of BP neural network, then using the resilient propagation (Rprop) training algorithm trained the BP. Experiments demonstrate that not only the classifier can recognize the hand-sketched primitives of arbitrary directions and positions, but also its abilities of anti-noising and identifying are robust, and the identifying rate is high. Furthermore, the classifier needn't be trained again in application.