| |
| This paper proposed a novel descriptor, called multi-scale triangle measurement (MSTM) for shape retrieval. The original shape is represented by a serious of uniform sample points, and each sample point is described by geometric measurement (rotational angle and side lengths) in different scale levels based on some triangles that consist of this sample point and its adjacent critical points, which are abstracted with polygonal approximation approaches. The generated triangles investigate the local variance (e.g. local deformation) and global information (e.g. topology) among different scale levels. After computing MSTM, the usual dynamic programming (DP) technique is employed based on the uniform sample points. The novel descriptor was applied on two well-known datasets: MPEG-7 and Tari1000. Experiments show that our descriptor achieves a retrieval rate comparable to state-of-the-art on the MPEG-7 data set, and outperforms other algorithms on the Tari1000 data set. Our descriptor has less consumption in feature computing than certain classical descriptors, e.g., the Inner Distance Shape Context (IDSC). |
|
| Keywords:image retrieval; shape matching; contour evolution; multi-scale analysis |
|
