Fast template matching algorithm for line pictures based on its chain coded description

Abstract

In this paper, a new matching algorithm for line pictures is presented. In the face authentication system developed in our laboratory, the isodensity lines has been introduced to differentiate between the facial features that are relevant for identification; this technique have a high level of discriminative accuracy. However, there still remain some problems to be improved in this system. In the conventional system, image based matching algorithms have been adopted. Therefore, the matching performance must be improved in terms of processing time and memory requirements. The isodensity lines are the line pictures extracted from a face as the contour lines of a face representing its relief. Therefore, these line pictures can be easily transformed into chain codes. By using these coded isodensity lines, remarkable improvement in the processing performance can be expected in terms of the processing time and memory requirements. However, even if the chain code can be adopted, there still are problems with transformation and the matching algorithm for line pictures. 1. The isodensity lines are very complicated in shape, therefore some new techniques must be introduced for transformation and matching. 2. In the matching stage, the similarity between two isodensity lines come to be extremely different from each other compared with the ideal case even if the differences between two chain codes are very small in length and direction. From this viewpoint, the matching process based on the chain code become very difficult in general. To resolve these problems, a new algorithm must be devised. In this paper, a new matching algorithm between two line pictures described by the chain code is presented and the efficacy of the proposed method is also discussed along with the experimental results. From the computer simulation performed using images of 50 people, it turned out clear that the processing time was decreased to approximately one-seventh (15%) compared to the conventional method. With respect to memory requirement, it was reduced to a quarter