In order to realize accurate measurement of aircraft’s current attitude, how to improve real time and robustness of star pattern recognition is the key of star sensor. The algorithms for star pattern abstraction, training sample set creation and network training improvement are proposed. First, a method of triangulation based on the character of star image is designed to combine all the stars of current field of view, which is used to extract star pattern and create complete training samples. The character of star pattern extracted has the advantages of translation and rotation invariance. Then BP Neural Network serves to recognize the star pattern with the weight matrix instead of navigation library. It is very fast to acquire current star information when the network has finished training. Particle Swarm Optimization (PSO) serves to train BP Neural Network, which helps BP network converge to the most optimum value. The experimental results show that the success rate of accurate recognition is 100%.