|dc.description.abstract||As the computer science progressing, image processing based droplet sorting system has been reported. Image processing is the most direct method of observation. It can detect target by target contour, size, shape, transmittance, etc. It is helpful for biological research because cells can be separated without any treatment. In this thesis, we integrate an image classification system with (support vector machine) SVM machine learning and an electric droplet sorter to try to develop a new droplet sorting technology, label-free image sorting system.
In this system, droplets are formed and naturally encapsulate micro particles. Then, continuous fluid is injected to fit system requirement by controlling droplet condition like droplet separation. Finally, the system sort droplets with particles out.
To further study this system, we divide the droplet sorting system into 3 stages. In the first stage, we aim to stably generate monodisperse droplets and observe some critical parameters like droplet generation frequency, and droplet separation. From this experiment, we get the pressure combination that can be used in the third stage, droplet sorting system.
The second stage is to observe the encapsulation of 7 μm particles. In the experiment, we verify its physical principle and get an estimate equation (eq. 4.2.4). This equation estimates the probability of occurrence of droplets containing different numbers of particles and help for getting droplet with single particle.
In the third stage, we study a droplet sorting system. For SVM machine learning image detection system, we first train this classification model and verify its accuracy. If its learning process is successful, the system is integrated to sorter and test system accuracy. From our experiment, we believe our system already has ability to identify droplet with or without particles and classify at a rate of 50 droplets per second. We also find the system accuracy and sorting accuracy are up to 95% and 100%, respectively. In addition, we are training the multi-classification model to find out single particle droplets. The system accuracy has reached 97.5% and 97.73% depends on strategy.