This paper addresses the problem of detecting small targets with very small spatial extent that are masked by spatially strong background clutter in image data. A novel adaptive prediction algorithm of background clutter is proposed based on decreasing the Cumulation-squared Error (CSE) of gradient in neighborhood. The algorithm can improve the Signal Clutter Noise Ratio (SCNR) of small targets evidently. Many experiments prove that the algorithm has better performance than traditional methods. And a fast detecting algorithm based on statistical analysis is introduced. It can get higher probability of detection based on lower probability of false alarm than other methods. The theoretic analysis and many simulations verify the validity of the method.