Research On Automatic Recognition Technology Of Mechanical Watch Reading Based On Vision

According to the detailed rules for evaluation of mechanical watches,high-precision mechanical watches can only be certified when they can meet the standard of High-precision Mechanical Watch SQL / HSTU 009-2017.In order to produce qualified mechanical watches,travel accuracy measurement experiments will be conducted by the watch industry before these watches leave the factory.The experiment observes the deviation of the reading of mechanical watche s compared to the standard time so that the watch industry can recalibrate the unqualified watches.The manual reading method used in the traditional mechanical watch reading detection is inefficient and easy to be affected by the status of the staff.In order to increase the output of qualified mechanical watches and adapt to the rapidly increasing market demand,it is required to design an automatic method to measure the reading of mechanical watches.1)In the research field of automatic reading recognition technology,there are many reading recognition technologies for instruments,but few for mechanical watches.Owing to the difficulties of the identification technology,such as the complicated dial patterns,numerous pointers,and the hig h possibility of pointer occlusion and so on,this paper proposes a vision-based automatic reading recognition method for mechanical watches.The research content can be divided into several parts including dial area extraction,pointer and scale area extraction,reading recognition and system design.The details are as follows: For all kinds of mechanical watches under various backgrounds,an image pyramid template matching algorithm combined with edge information is used to extract the dial area.Compared with other existing extraction methods that cause problems such as background interference and distortion,the algorithm in this paper has higher robustness.2)Aiming at the problems of troubling in recognition resulted from the difficultly of parameter setting,unclear segmentation,pointer occlusion and multiple scales while adopting the traditional method to re cognize the pointers and scales,an improved Mask R-CNN network combined with context information is proposed,which can adapt to the extraction of pointer and scale under various color jitter and noise interference.At the same time,a data set of mechanical watch pointers and scales is created,which can be used for the training of deep learning networks.3)As for the reading recognition,the traditional identification method of pointer axis needs to rely on the effectiveness of algorithm refinement.To solve the problem,the paper proposes an improved least square distance method,which can accurately locate the pointer axis.Meanwhile,making use of the least square method combined with the prior information,the inner circle of the dial is fitted and the center of the dial is determined.According to the angle relationship between the central axis of each pointer and the connecting line between the dial center and the scale center,the reading of the mechanical watch is calculated.In order to further improve the accuracy of mechanical watch reading recognition,the principle of error correction is put forward according to the implied angle relationship between each pointer.4)Based on the automatic recognition scheme of mechanical watch proposed in this paper,a software system for automatic reading measurement of mechanical watches is designed,including the design of interface and database,and then an error analysis of the experimental results of complex mechanical watches is carried out which shows that the reading error is w ithin 1s.At the same time,the accuracy measurement of mechanical watch is realized by using this a lgorithm,which can improve the measurement efficiency.