고속 처리용 문자 특징의 성능 비교

Abstract

DSP 소자등을 사용하여 문자 인식 시스템을 구성하는 경우에 연산량, 인식률의 측면에서 사용 가능한 특징들의 성능을 비교하여 보았다. 비교된 특징들은 메쉬(mesh)와 receptive field를 달리하는 몇 종류의 방향성 에지 연산자들이다.

이 특징들을 문자 인식에 적용하여 본 결과, 메쉬 보다는 방향성 에지 연산자들의 인식률이 높았으며, receptive field를 작게 하여 특징 벡터의 크기를 크게한 것이 인식률이 좋아지지만 연산량은 늘어나게 되어 시스템 구성에 부담이 된다. 본 논문에서 선택한 적절한 특징은 수평 방향의 receptive field, 수직 연산자에 수직 방향의 receptive field를 적용한 것으로서, 이들은 벡터 크기와 인식률에서 적절한 것으로 판단된다.

The performance of mesh and directional edge operator with different receptive fields, that are thought acceptable in the recognition rate and tin the amount of arithmetic operation, are compared. These features are for high-speed character recognition systems based on DSP components.

The results we got from the experiments of character recognition are as follows. The recognition rate of the directional edge operator is superior to that of the mesh. The directional edge operator with small receptive field explores better recognition rate than the operators with other receptive fields, but the vector size of the former becomes twice of the latter. It means the increasing of the amount of arithmetic operation. As a result, the appropriate feature among the features we tested for high speed character recognition is the mixed feature of horizontal and vertical operator with horizontal and vertical receptive field, respectively.

The performance of mesh and directional edge operator with different receptive fields, that are thought acceptable in the recognition rate and tin the amount of arithmetic operation, are compared. These features are for high-speed character recognition systems based on DSP components.

The results we got from the experiments of character recognition are as follows. The recognition rate of the directional edge operator is superior to that of the mesh. The directional edge operator with small receptive field explores better recognition rate than the operators with other receptive fields, but the vector size of the former becomes twice of the latter. It means the increasing of the amount of arithmetic operation. As a result, the appropriate feature among the features we tested for high speed character recognition is the mixed feature of horizontal and vertical operator with horizontal and vertical receptive field, respectively.