In the complex environment of clutter edges and multiple targets, it is the key to improve the capability of CFAR processing to establish a stable adaptive detection technology. A method of CA/OS-CFAR detection based on neural network is proposed based on cell averaging-constant false alarm rate and cell averaging/ordered statistic-constant false alarm rate. Use the neural network to determine the optimal detection method, according to the selected detection method to calculate the detection threshold to improve the ability of radar detection target. The input of the neural network contains CA, OS-CFAR and the unit value to be measured. The initial threshold is calculated by training, and the type of input is classified and recognized by neural network. The threshold is compared with the results of CA-CFAR and OS-CFAR, and the optimal threshold is selected. This method is tested with a simulation case of homogeneous clutter, multiple targets and clutter edge environment. Experiments show that the method proposed in this paper can be used to determine the optimal detection method in the mean and non uniform clutter background.