基于轻量化特征选择的镜片多尺度缺陷检测系统研究  

Research on Multi-Scale Defect Detection System of Lens Based on Lightweight Feature Selection

作　　者：许桢英[1] 杨钰峂 雷英俊 王匀[1] 武子乾 韩丽玲 XU Zhenying;YANG Yutong;LEI Yingjun;WANG Yun;WU Ziqian;HAN Liling(School of Mechanical Engineering,Jiangsu University,Zhenjiang,Jiangsu 212013,China)

机构地区：[1]江苏大学机械工程学院,江苏镇江212013

出　　处：《计量学报》2024年第6期786-794,共9页Acta Metrologica Sinica

基　　金：国家自然科学基金(51679112)。

摘　　要：为解决对镜片缺陷识别和定位困难的问题,设计了一种基于轻量化特征选择(LFSN)的镜片多尺度缺陷检测系统。首先通过四步相移栅格光成像系统采集缺陷图像,基于傅里叶变换对图像进行融合以提高图像质量;然后,LFSN使用自动特征层选择结构,在训练过程中计算无锚框分支损失,获得最优特征层信息,更新参数,从而优化模型对不同缺陷大小信息的学习能力;还使用了深度分离可变卷积,通过双线性插值增加像素点在平面的偏移量,从而提升模型对缺陷形貌的主动学习能力,并一定程度减少模型训练参数量,降低检测时间,同时优化回归定位损失明确各阶段训练任务,利用一次惩罚项指导前期预测框中心距离回归;利用归一化二次项,指导后期预测框大小比例回归,使预测框更接近真实值;最后,通过实验采集镜片缺陷图像,并构建数据集进行对比实验。实验结果表明:识别和定位镜片的缺陷的准确率为96.3%,单幅检测时间为24.9 ms。In order to solve the problem of difficulty in identifying and locating lens defects,a multi-scale lens defect detection system based on lightweight feature selection(LFSN)is designed.To solve the problem that existing image acquisition,processing,and recognition methods are difficult to identify and locate defects in lenses,a multi-scale defect detection system for lenses based on Lightweight Feature Selection Network(LFSN)is designed.First,in order to increase the image quality,the fault images are fused using the Fourier transform after being obtained through the design of a four-step phase-shifted raster optical imaging system;Then,aiming to optimise the model s capacity to learn about various defect sizes,the LFSN computes the anchorless frame branching loss during training in order to acquire the ideal feature layer information and update the parameters;Additionally,the system employs depth-separated variable convolution to enhance the offset of pixel points in the plane via bilinear interpolation,thereby enhancing the defect morphology model s capacity for active learning and somewhat lowering the number of model training parameters to shorten the detection time.Simultaneously,the optimisation of the regression localization loss identifies the training tasks in each stage.The early stage of the prediction frame s regression is guided by the primary penalty term,and the late stage of the prediction frame s regression is guided by the normalised quadratic term,which brings the prediction frame closer to the real value.Lastly,a dataset is created for comparative studies and lens defect photos were obtained experimentally.The research data demonstrate that this inspection method has a 96.3%accuracy of identifying and locating defects in lens,with a single frame detection time of 24.9 ms.

关 键 词：光学计量 镜片 缺陷检测 栅格光成像 轻量化特征选择 混合交并比 

分 类 号：TB96[机械工程—光学工程]