Abstract: Interference noising originating from the ultrasonic testing defect signal seriously influences the accuracy of the signal extraction and defect location. A fast algorithm of ultrasonic defect extraction is proposed which combines genetic algorithm with sparse decomposition together in this paper. The genetic algorithm is employed to optimize multi-Parameter of sparse decomposition. According to the characteristic of ultrasonic echo, the best Gabor atom is chosen to match self-adapting the ultrasonic defect signal. Therefore it can greatly reduce complexity of sparse decomposition. Both simulated and actual steel defect signals were tested, and compared with the wavelet transform, and experimental results showed that the signal could be then reconstructed efficiently from noise.