To address the difficulty in detecting a dim-small target in single frame image caused by the complex background and polymorphism of the target, a method of rough extraction for threshold segmentation and precise detection for multi-point signal-to-noise ratio (SNR) is proposed. In the rough extraction stage, an improved threshold segmentation algorithm based on robust principal component analysis (RPCA) is proposed. The ratio of the mean value of the neighborhood sparseness to the mean value of the whole sparse image is used for the threshold segmentation, so as to further eliminate the isolated noise and the edge clutter of background cloud. In the precise detection stage, a multi-point constant false alarm detection algorithm based on statistical characteristics is proposed. The SNR of each pixel of candidate points in the neighborhood is obtained, and then the target point is extracted based on the false alarm rate threshold and statistical quantity threshold. The problem of polymorphic features caused by the dispersion of target energy will be overcome. Experimental results show that the detection probability of this algorithm reaches 95.6% under complex background, and the false alarm rate is 56.1% and 47.1% lower than that of single pixel and neighboring pixel based SNR computing methods, respectively.