Research On Scratch Detection Of Optical Lens Surface Based On U-Net Network

Optical lens surface scratch is the most common product defect in lens production,and the diffraction and scattering phenomenon caused by scratch will seriously affect the optical performance of lens.Whether the scratch information can be accurately detected will directly affect the rating of lens quality,which has always been the concern of lens manufacturing industry.At present,the defect detection method of lens manufacturers is mainly manual visual inspection,which is often inefficient and unable to carry out qualitative and quantitative analysis of scratch defects;Domestic researchers generally focus on the basic classification of lens defects.However,due to the limitation of algorithm performance,the specific parameters of defects are mostly estimated and measured,so the accuracy can not reach the enterprise rating standard,and the scratch and similar defects can not be accurately distinguished.Therefore,this paper proposes a new method based on U-Net network for optical lens surface scratch detection:First of all,through the analysis of optical imaging principle of lens surface defects,an image acquisition system based on coaxial light source is designed.The light source,light color,camera parameters and lighting mode are debugged.The problem that the image acquisition is affected by the strong specular reflection of the transparent surface is solved,and the high-quality lens image is obtained.Secondly,in order to solve the problems such as incomplete main body,unclear edge,unable to accurately segment with background and other defects caused by traditional image segmentation algorithm,combined with the characteristics of deep learning and high-precision detection requirements of scratches,U-Net deep learning network model is selected to realize automatic segmentation of scratch image.However,due to the main body deformation and fuzzy boundary in the original network image segmentation results,after analyzing the reasons,we improved the original network by adding BN layer and expanding convolution structure to achieve high-precision segmentation of optical lens scratch image.On this basis,through the analysis and comparison of a variety of algorithms to achieve the number of scratches,location,length and width measurement,and design an interactive interface to facilitate the staff to operate the detection system.Finally,the above algorithm is verified by experiments.The results show that the scratch segmentation accuracy based on the improved U-Net network model reaches more than 99%,the false detection rate of scratch defects is less than 1.5%,and the measurement accuracy reaches 0.03 mm.This method realizes the qualitative and quantitative measurement of scratch defects,and provides a new method and idea for lens manufacturers to detect the scratch on lens surface and grade the lens quality.