Contour-based Moment Invariants and Their Application to the Recognition of Object Shapes
轮廓矩不变量及其在物体形状识别中的应用

Object recognition is a challenging problem in the field of pattern recognition and computer vision. Hu's moments are classical tool in the field, which are defined based on the colors or gray levels of objects. This paper is an improvement of Hu's moments. A series of novel moments, which are called contour moments, are constructed based on object contours and their applications to object shape recognition are given in this paper. Some properties of these new moments including the invariance on shift, rotation and scale transforms are studied and proved. A central advantage of the new moments over Hu's moments is that they are independent of the colors or gray levels of objects. They are defined completely by the contours of objects, namely, that they are completely the shape features of objects. To support our new theory, an algorithm for object shape recognition is designed based on the new moments and experiments are conducted. In our experiments, wavelet transforms are employed to extract the contours of objects, therefore, a brief introduction on the theory of wavelet transform as a multi scale edge detector is introduced. Considering that an object may have more than one contour, each of which is a close curve, this paper also gives detailed discussion on how to deal with several contours. Experiments give an encouraging high recognition rates.