Research On Inspection Of Bearing Size Based On Deep Learning

The result of the workpiece size inspection is an important indicator that determines whether the quality of the product meets the production requirements and conforms to the factory standard.For traditional workpiece measurement,it is necessary to clean the workpiece first before performing manual measurement.For manual measurement,there are some shortcomings,such as low efficiency and missed inspection.However,machine vision has the advantages of high efficiency,wide applicability and safety,which solves the problems faced by manual measurement and has high research value.This thesis is a study on the online size detection of bearings based on deep learning,which mainly focuses on the interference area detection,interference area processing,edge detection and verification through experiments.The main research contents are as follows:In the interference area detection stage,it mainly deals with the interference areas on the surface of bearings in bearing production.Semantic segmentation of the interference area is implemented and an improved fully convolutional neural network is proposed to detect the interference area of the detected bearing images.Then,the semantic segmentation results of the improved method and the fully convolutional neural network are compared.In the interference area processing stage,it mainly focuses on image restoration of the detected interference areas.Color difference of the semantic segmentation images are used to cut the interference areas so as to obtain the binary mask of areas that need to be repaired,and areas that need to be repaired are marked through the mask.Then,the depthwise separable convolution is used to achieve the lightweight of the image restoration model based on the context of the contextual attention and improve the speed of image restoration.In the edge detection stage,in view of the limited accuracy of pixel-based edge detection methods,an edge detection method combining pixels and sub-pixels is adopted.By comparing the effects of traditional operator edge detection methods,rough detection results of workpiece edges turn out to be better.So as to perfect the detection results accuracy,three sub-pixel edge detection methods,the cubic spline interpolation method,the gray scale moment method and the polynomial fitting method,were all compared,and then the method with the best accuracy was chosen to increase the detection of accuracy to the sub-pixel level.In this thesis,various experiments are constructed to check the above method’s results.Experiments have proved that the semantic segmentation method proposed in this thesis is able to effectively detect the interference area.Also,this approach improves a lot in semantic segmentation effects and other indicators in comparison to the fully convolutional neural network.The size detection can also be well completed based on the method of sub-pixel edge detection.Meanwhile,the measurement accuracy of the sub-pixel method of gray moment is about 0.02 mm,which conforms to the bearing production standards.