Aiming at the problem of the classification of musical instruments

based on concept of nonlinear dynamics

the audio time series generated by different instruments are processed by the recurrence plot method

each set of audio time series corresponds to a recurrence plot

the dense scale-invariant feature transform（ SIFT）feature of each recurrence plot is extracted. Dimensionality of high-dimensional SIFT features by bag-of-visualwords model

the frequency histogram vector of each recurrence plot is obtained as a feature input linear support vector machine. One-versus-rest of support vector machine（ SVM） algorithm is utilized to achieve multiclassification. Experiments show that the recognition accuracy of this method is higher than the recurrence plot dense ratio feature method

and it is 3 %～ 4 % higher than the Mel frequency cepstrum coefficient method

which provide new ideas for instrument recognition.