Font Size: a A A

Study On Droplets Segmentation And Characteristic Detection Of Microfluidics Based On Deep Learning

Posted on:2022-01-08Degree:MasterType:Thesis
Country:ChinaCandidate:M ZhangFull Text:PDF
GTID:2518306323967109Subject:Instrument Science and Technology
Abstract/Summary:PDF Full Text Request
Microfluidic droplets can be used as a micro reactor for biochemical reactions,and have a wide range of applications in the fields of biology,chemistry,medicine et al.In specific application,it is usually necessary to accurately measure the size and quality of the single droplet.Recently,deep learning is usually used in microfluidic measurement,such as the measurement of fluid velocity and concentration by using microfluidic droplet image.By analyzing the basic flow pattern of T-shaped microchannel,it is found that the formed droplets and the squeezing-dripping regime during the formation process have different characteristics.On the other hand,machine learning methods are not universal for the identification of the dispersed phase region such as the squeezing-dripping regime and droplets under different flow patterns.Therefore,this paper aims to develop a semantic segmentation network suitable for microfluidic droplet measurement,which can accurately capture the dispersed phase regions such as droplets and the squeezing-dripping regime in different flow patterns.Firstly,the traditional image segmentation algorithm,deep learning technology and deep learning computer vision technology are investigated.By adding skip connections,padding operation and batch normalization operation,a semantic segmentation network with three layers of input and output,from image to image and three layers of small modules for decoding and coding is constructed.Three different loss functions and two different optimizers are trained to get the most suitable network model.By comparing the network test results with different loss functions and optimizers,it is found that Adam optimizer achieves better results in the network iterative process.It is also found that Dice loss can solve the problem of unbalanced data in the network of microfluidic droplet segmentation,and improve the precision of segmentation network effectively.It shows that using Dice loss and Adam optimizer,microfluidic droplet semantic segmentation network obtains good results.The deep learning method,which is proposed in this paper,is used to detect the micro fluid droplets in the T-type microchannel using an image at a certain time.The original droplet image is input into the semantic segmentation network to get the segmentation image.After image processing of each semantic region,the important parameters are obtained successfully,including the center position of droplets,size(including area,width and height),velocity of the droplet,as well as the continuous phase liquid film thickness of the squeezing-dripping regime when the droplet is being generated.The important parameters during droplet formation,such as the width and height of the droplet head,are obtained perfectly.The experimental study of droplet generation was carried out on the T-shaped microfluidic channel,by changing the continuous phase flow rate Qc and the flow rate ratio q of dispersed phase to continuous phase.The flow patterns of dropping,squeezing and parallel flow were used to observe and photograph.Semantic segmentation of droplet images under different experimental conditions was carried out,and the characteristics of each region were detected and analyzed.At the same continuous phase flow rate,the droplet size increases with the increase of flow ratio,and the detected droplet size aggregation degree is good.In addition,the droplet height is stable numerically under different flow ratios,which proves the stability of the droplet detection method in this paper.Through the analysis of the droplet center position and the characteristics of the squeezing-dripping regime under the same experimental conditions,the characteristics of each droplet change according to the law under the same period,which is consistent with the actual observation.
Keywords/Search Tags:Microfluidics, Droplets, Deep learning, Semantic segmentation, Characteristic detection
PDF Full Text Request
Related items