Fiber positioning technology is widely used in spectroscopic telescopes, and the accurate identification of the fiber position on the focal plane directly affects the efficiency of the astronomical spectrum. At present, fiber positioning usually uses the “back-illuminate” technique to illuminate one end of the fiber. The other end of the fiber is used for detection. The fiber could be stressed or twisted during locator motion, resulting in a difference between the detected fiber position and the actual fiber core. However, the fiber-optic back-illuminated device in the spectrometer system increases the complexity of the system and the time loss of fiber positioning. This paper attempts to use a new method combining image processing with deep learning to identify the fiber ferrule by the front-illuminated method. We built an experimental platform in the lab and experimented with a CMOS camera and telecentric lens. We tested the repeated errors and displacement measurement errors of the two methods. A series of comparative experimental results show that the final detection accuracy of this method can meet the needs of optical fiber positioning in the laboratory, although it has not yet reached the accuracy of the back-illuminated approach. In the future, if the light source and fiber ferrule were specifically designed for the front-illuminated method, its accuracy could be further improved.