Efficient contour-based shape representation and matching

This paper present s an e#cient met hod for calculat ingt he similarit y bet ween 2D closed shape cont ours. The proposed algorit hm is invariant t ot ranslat ion, scale change and rot at ion. It can be used for dat abase ret rieval or for det ect ing regions wit h a part icular shape in video sequences. The proposed algorit hm is suit able for real-t ime applicat ions. Int he first st age oft he algorit hm, an ordered sequence of cont our point approximat ingt he shapes is ext6 ct d fromt he input binary images. The cont ours aret ranslat ion and scale-size normalized, and small set s oft hemost likely st art ing point s for bot h shapes areext ract ed. Int he second st age,t hest art - ing point s from bot h shapes are assigned int o pairs and rot t on alignment is performed. The dissimilarit y measure is based ont he geomet rical dist nces bet ween corresponding cont our point s. A fast sub-opt imal met hod for solvingt he correspondence problem bet ween cont our point s fromt wo shapes is proposed. The dissimilarit y measure is calculat ed for each pair of st art ing point s. The lowest dissimilarit y is t ken ast he final dissimilarit y measure bet weent wo shapes. Three di#erent experiment s are carriedout usingtn proposed approach: letR r recognit ion using a web camera, our own simulat ion of Part B oft he MPEG-7 core experiment CE-Shape1 and det ect ion of charact ers incart oon video sequences. Result s indicat et hat t he proposed dissimilarit y measure is aligned wit h human int uit ion.