Abstract: A method of circular marker detection in under-exposed helicopter blade images based on YOLOv3 (you only look once) and watershed algorithm is proposed, aiming to improve detection adaptability, speed up the detection and obtain accurate position of circular markers. Firstly, the real under-exposed blade images are labeled for dataset making, on which the YOLOv3 network is trained. Secondly, the circular marker regions in blade images from testing dataset are detected by the trained YOLOv3 network. Thirdly, the traditional watershed marker detection method is improved, and the circular marker regions are used for the watershed transformation by the multi-threading technology in parallel, and the edge detection results of the circular marker are obtained. Finally, the circular markers are accurately located by the least square circle fitting and the method of removing the singular points. The proposed method is proved to be adaptable, fast and accurate by a number of experiments in many under-exposed helicopter blade images, and has been applied to circular marker detection in under-exposed helicopter blade images.

Key words: under-exposed image, circular marker detection, YOLOv3, watershed transformation