To detect dim and small target in infrared image with low signal-to-noise ratio, a novel algorithm based on multi-scale gradient of gray level is presented. Firstly, for all points in an image, the algorithm selects up, down, left and right reference points around the current point according to the size of target, and identifies potential targets based on the maximum gradients between current point and reference points. Then, final targets are identified with the fusion of potential target detection results of the three sequential images. The algorithm does not need background prediction, and achieves not only higher detection rate but also lower false alarm rate in infrared image with low signal-to-noise ratio and complex background. Experimental results show the validity of the algorithm.