Research On Defect Detection Technology Of Ceramic Tooth Zirconium Disc Based On Deep Learning

Ceramic tooth is a kind of more popular and cost-effective dental restoration.In the production process,the inspection of the surface quality of the ceramic tooth zirconia disc is very important.In this paper,a system for surface defect detection of ceramic tooth zirconium discs based on deep learning is designed,which can effectively identify and classify target defects,and design a visual interface to display the detection results.The main contents are as follows:(1)The overall design plan of the entire defect detection system is described,and the construction plan of the hardware module and the software module is given.The construction of the hardware module includes the selection of industrial cameras,lenses and light sources in the acquisition system,as well as the construction of a complete defect detection device combined with the transmission system.The construction of software modules includes data set production,YOLOv3 defect detection network construction and network optimization,and visual interface design.(2)Aiming at the problem of defect detection on the surface of ceramic tooth zirconium discs,a defect detection network based on deep learning was built.First,for the insufficient number of samples,data enhancement methods such as cropping,flipping,and rotation are used to expand the data.Then use Labellmg software to annotate the pictures,and complete the data set production according to the format of VOC2007.Next,the YOLOv3 target detection network and the YOLOv3-Tiny target detection network are introduced.For small target object detection,the YOLOv3 defect detection model is improved: the backbone network is lightweight,the Inception v3 structure is introduced,the dimensional clustering candidate frame is improved,and the design of the target loss function.The improved YOLOv3 algorithm has improved detection speed and accuracy.(3)The target detection algorithm based on YOLOv3-Tiny has the advantages of fast detection speed,small size,and easy deployment on edge devices.At the same time,it also has the problems of low detection accuracy and inaccurate positioning of small targets.This algorithm is improved.First,improve the network structure,design a new backbone network while ensuring real-time performance,and improve the feature extraction capability of the network.Secondly,the target loss function and the feature fusion strategy are improved,and the IOU loss function is used to replace the original frame position loss function to improve the positioning accuracy.The improved YOLOv3-Tiny algorithm in the VOC-Z test set increased the m AP by 5.6%,the reasoning time only increased by 3.3ms,and the detection performance was further improved.(4)Through the actual application on the production line of dental restoration products,the model recognition effect is tested.The results show that the recognition rate has reached the expected result(91.6%),and it is displayed on the visual interface,which meets the requirements for the surface quality inspection of ceramic tooth zirconium discs in industrial production.