Automatic bubble level correct system based on deep learning
                                                                            LIU Yao, ZHU Shan an
                                   （College of Electrical Engineering, Zhejiang University, Hangzhou 310027, China）

Abstract: Aiming at the automatic correction of the bubble level with large angle of inclination, the models in the field of deep learning object detection were studied, and a method based on deep learning object detection models was designed. The YOLO and SSD models were trained on the data set of bubble level images with different light conditions. By adopting K means clustering algorithm, the effect of different anchor box numbers on YOLO model s performance was analyzed. And two models  detection accuracy and average IOU werecompared. Two edge fitting methods combining Progressive Probabilistic Hough transform with least square method were designed and two methods  accuracy was compared. Aiming at this application, Client/Server network structure was used. The images were sent to the server and calculated.The calculation results were sent back to the client to control motor to correct the bubble level. The results indicate that the proposed method can detect the bubble level with large angle of inclination, and is more efficient than the conventional methods.

Key words: bubble level; YOLO; SSD; progressive probabilistic hough transform; least square method