Aiming at the problem that the polarization axis detection method based on end surface is not robust to light and can not detect polarization axes of several polarization maintaining fibers at the same time, a detection method based on Faster RCNN was proposed. Firstly, the Faster RCNN neural network model was trained and the robustness of stress rods recognition was enhanced by parameter tuning. Then, the edge points of stress rods were detected by the subpixel edge detector based on Zernike moment to improve the measurement accuracy. Lastly, the relationship between measurement accuracy and error is analyzed. Experimental results show that the measurement accuracy of the proposed method is less than ±0.1°, the robustness of light is enhanced and the detection of polarization axes of multiple polarization maintaining fibers is realized at the same time.