Because of the complicated mechanisms of atmospheric turbulence, it is difficult for existing image restoration algorithms to achieve a good balance between restoration efficiency and quality. A technique for the acceleration of support vector machine based image restoration algorithm was presented to solve this problem. In this algorithm, variance threshold was applied to assist sample selection. Since redundant information was discarded by means of the sample selection algo- rithm, training samples became more effective. Meanwhile, a real-time model updating algorithm was applied to speed up the restoration rate of serial images. Comparisons of the acceleration method with the previously proposed restoration algo-rithm on electric arc wind-tunnel images were provided. Experimental results show that the proposed acceleration method runs faster and performs better. It can effectively restore the turbulence-degraded images from strong noise, energy attenuation, blur and jitter.