In order to solve the problem of detecting moving small targets in low-contrast infrared image sequences, a new detection method based on multilevel filter is proposed. First of all, for small targets which had certain space distribution scope(≤5×5), frequency characters of target, noise and background were analyzed in detail in terms of the gray distribution model of target, noise and background. Therefore, the conclusion was got. Background was relatively dominant in low frequency part, target was viewable in middle frequency part and noise was more obvious in high frequency part in the frequency domain in terms of Fourier transform. The frequency scope of various sizes of target could be estimated by the Fourier transform of GATE function. Then various sizes of candidate targets could be distinguished from complex background by different level filters. Real targets could be detected by properly processing these images which contain targets.