In this research, a new method of time-frequency image feature extraction is proposed. It is based on the multi-scale and multi-direction decomposition of the non-subsampled contourlet transform. Firstly, the vibration signals are transformed into a time-frequency image which is then converted into a gray image based on the contourlet transform. Then, the coefficients at high and low frequency are calculated based on the gray image. Different feature extraction methods are investigated in details. In this paper, the energy at the high-frequency, the mean and standard deviation at the low-frequency are calculated as characteristic parameters. Finally, the data of different conditions from a gearbox and rolling bearing are classified and tested by support vector machine (SVM). The results show that the proposed method is effective in determining the characteristic value of a time-frequency image for the condition identification.