A new time-frequency optimal feature extraction method based on wavelet packet decomposition and sequential forward floating selection (SFFS) algorithm was presented in order to improve the correct rate and generalization in defects recognition through ultrasonic inspection. This method composes supervised and unsupervised feature extraction technologies to achieve better recognition performance by partial using prior classificatory information of samples. Theories evolved in the new method were introduced, including wavelet packet transform, Fisher criterion and SFFS algorithm. In order to verify the effectiveness of the new method in defect recognition, tests were carried out with four kinds of typical defects on oil casing pipes. Three traditional feature extraction methods from time domain, frequency and wavelet packet domain were employed respectively for performance comparison,and all the feature sets obtained by different methods were classified with support machine algorithm. The experimental results of 10 groups of random selected samples show that the new method can identify the defects effectively with a best recognition rate of 93.3% and an average rate of 89.5%.Compared to the three traditional methods mentioned above, the new method can achieve higher recognition rate and better generalization performance and is less sensitive to the selection of training samples.