Research On Image Detection Method Of Surface Defects Of Microfluidic Chip

Microfluidic chips integrate detection,sorting and other operations in biological and chemical laboratory into micro and nano scale chips,and are widely used in many fields such as medical treatment and scientific research.However,the dust,structural connectivity and other defects on the surface of microfluidic chips may lead to wrong experimental or diagnostic results.Therefore,the research on the surface defect detection method of microfluidic chips has important application value for improving the chip quality and avoiding the interference of defects.There have been a lot of related researches on the surface defect detection of microfluidic chips or micro-nano structures,and image detection technology has been widely used.However,most of the existing image detection technologies are applicable to defects of similar size.Microfluidic chip surface defects have a larger aspect ratio range and size range.When these methods are applied to chip defect detection,the results are not satisfactory.In addition,some unknown defects may appear due to softer chip materials and irregular processing.The existing methods are not ideal for identifying unknown structural defects.At the same time,the influence of external environment on image quality also increases the difficulty of defect detection.In view of the above problems,we use image processing,deep learning and other technologies to process and analyze the chip surface image to achieve accurate defect detection.The specific works are as follows:Firstly,in order to solve the problems of high similarity and misalignment of chip images,the accurate image registration was achieved through the difference of brightness caused by transmission imaging and chip structure,and the registration speed was improved by using rectangular edges as the matching template,and the registration time was reduced from 1279ms to 202ms.Through MASK dodging method and histogram specification,the uneven illumination is eliminated,and the details covered by the shadow are enhanced,effectively avoiding the influence of the shadow on defect detection.Secondly,we use Yolact++network to detect known defects on the chip surface,and design anchors based on differential evolution algorithm.And we enhance the learning of slender defects through Focal optimization objective.Different anchors were designed for different levels of FPN,and better detection effect was achieved with fewer parameters,especially for slender defects such as fiber and scratch,the effect was improved by 11.36%.The AP50 of defect prediction reached 93.59%.Finally,a multi-task learning method is proposed for the detection of unknown defects.The multi-task network improves the descriptiveness and compactness of the extracted features through two branches,which can effectively distinguish abnormal and normal samples.For one class classification,we propose a uniform sampling method based on superpixel segmentation and clustering algorithms,which can effectively balance the ratio of samples with different structures on the chip,avoid the rare structures being mistakenly identified as defects.Through a multi-task network and classifier,the effective detection of unknown defects was realized,and the AUC reached 0.9381.In this paper,we use machine vision and deep learning to achieve the detection of known defects and unknown defects,which can provide guidance for subsequent chip processing and using,and avoid the interference of defects on the experiment and diagnosis.It has high application value.