Off-line recognition of a handwritten Chinese zither score

A Chinese zither score is different form a western staff. The Chinese zither score is handwritten, and is a combination of fingerings, scales, and several different types of notes. In this paper, we first construct pattern classes for fingerings and scales we frequently play. A specific segmentation method is derived in accordance with the zither score. After segmentation, all meaningful individuals can be found out and the weighted cross counting feature is used to extract features. A cascaded architecture of neural network with feature map (CANF) is proposed to obtain high recognition rates. The CANF cascades a supervised neural network trained by back propagation (BPNN) with an unsupervised neural network, Kohonen's self-organized feature map (SOFM). The SOFM can reduce the dimension of feature space and remove the redundancy of features in transformation such that the learning time of BPNN can be speeded up and the recognition rate can be improved. In our experiment, a real Chinese zither score is segmented, and the CANF shows a 100% perfect recognition rate.