Research On MEMS Defect Detection Algorithm Based On Polarization Imaging And Deep Learning

The rapid development of manufacturing industry has raised the importance placed on detecting structural and visual defects in products.Companies are seeking to replace the inefficient and imprecise manual inspection with machine vision.Deep learning-based detection approaches can achieve excellent accuracy and generalizability with a minimal requirement for imaging conditions or manual features.They are now widely utilized in academia and industry for defect detection.We conducted research on unsupervised anomaly detection algorithms,generative data augmentation techniques and supervised target detection algorithms with different stages of data quantity.The research objects were MEMS acoustic thin films.The main work of this paper is summarized as follows.1.In order to detect low illuminated MEMS defects while lacking defect samples,an unsupervised anomaly detection algorithm based on Do FP polarization image enhancement and Patch SVDD was proposed.Stokes vector images were used in Polarization demodulation and enhancement,which can expand the contrast between the defect and the background.The enhanced images were used to improve Patch SVDD in correctly generating heatmaps and segmenting defects for both normal and abnormal samples.The AUROC achieved 0.996,which was 22.4% higher than that of low illuminated images.2.MEMS acoustic thin film pictures were captured in an ultra-clean laboratory by an industrial CCD and a metallurgical microscope.Faced with an insufficient training set for target detection,640 training images were created by cropping via sliding window and training generative adversarial networks.The characteristics of DCGAN,LSGAN,WGAN,and several improved networks were analyzed theoretically and experimentally.The improved model WGAN-DIV-DC,which can generate images with lowest FID and most diversity,were used to make data augment.It expanded number of images in detection training set from 200 to 4000.3.The obtained 5000 MEMS defect images were labeled with classes and locations of defects using the Label Img under Python 3.9,and were divided into training set,validation set and testing set.The network structure of YOLOv3 and YOLOv5 are analyzed in terms of input,backbone,neck,and detection head.Improvements of YOLOv5 were introduced in detail,such as anchor boxes clustering,cross-stage connection,feature fusion and loss function.In defect detection experiment,the optimal model YOLOv5-aug was obtained by comparing the precision,recall,m AP and speed during training and testing.The optimal model was used to detect MEMS acoustic thin film defects,which achieved a m AP of 0.910 and a real-time speed of 79 fps.