In the visual robotics area, we should dope out a solution to detect small objects. We need to explore much more precise detection methods to achieve better performance. In this paper, a small object detection method based on the Relevance Vector Regression (RVR) is proposed. Compared to the Support Vector Regression (SVR), it avoids the set of free parameters and Mercer kernels. After regression process by RVR, the kernel correlation coefficient is used to increase the signal-to-clutter ratio (SCR). For improving the detection performance further, the post-processing method is applied after RVR and getting the nonlinear correlation coefficient. The experiment results prove that our method has the validity and better performance than the other two methods based on filter-based methods.