Investigation of microstructure movement under flow by using image processing and deep learning

thumbnail.default.alt
Tarih
2023
Yazarlar
Khosroshahi Sarbazzadeh, Saeed
Süreli Yayın başlığı
Süreli Yayın ISSN
Cilt Başlığı
Yayınevi
Graduate School
Özet
In many industrial and biological applications, the viscosity of chemical and biological fluids is a crucial material property that needs to be precisely measured A variety of techniques has been developed to measure viscosity. The micropillar-based microfluidic viscometer method uses viscosity and flow-induced micropillar displacement. Before making general viscosity measurements, calibration curves (viscosity vs. micropillar tip displacement) are created using solutions with known viscosities. Filming experiments with glycerol/water solutions with viscosities ranging from 2 to 100 cP at fixed flow (shear) rates are done to achieve this. The experiment is then repeated using a fluid sample whose viscosity is determined. In captured experiment videos with the sample fluid, the displacement of pillars is measured for this purpose using ImageJ, an image processing program. The measured displacements are then mapped to the calibration curves to determine the sample fluid's viscosity. Results obtained using this method are precise. The disadvantage is that using ImageJ to calculate displacement takes time and requires manual work. Therefore, in this study, we used two distinct image processing algorithms that yield results much more quickly. These are Lucas-Kanade (KLT), and Hough Circle, which are used in classical video processing, and the FlowNet2 neural network model. The KLT algorithm is a popular differential technique for estimating optical flow. We tracked the four corners of the pillar tip using the KLT to determine the displacement of the pillars. The final displacement data was then determined by averaging these four corner displacements. Contrarily, the convolutional neural network (CNN) Flownet2 is employed in deep learning to interpret visual images. Huge displacement control and precise estimation of minute details in the optical flow field are two features of FlowNet2. We made use of the pre-trained FlowNet2 model on the THINGS dataset. We used the first frame of the video as the model's first entry and the frames that came after that as its second entry to find out about displacement. We used ImageJ data as a reference to determine the methods' accuracy when determining the accuracy of the suggested methods for 10 videos. Regarding ImageJ, KLT, Hough Circle and FlowNet2 provided an average accuracy of 95.45%, 91.47%, and 95.62%, respectively. We saved a lot of time by using these techniques because we didn't need human assistance. With KLT, we were able to generate viscosity results 158 times faster with respect to ImageJ, with Hough Circle 396 times faster and with FlowNet2 10 times faster.
Açıklama
Thesis (M.Sc.) -- İstanbul Technical University, Graduate School, 2023
Anahtar kelimeler
microstructure movement, image processing, deep learning
Alıntı