Fast Image Segmentation Toward Automation of 3D Ice Printing

Abstract

Freeform 3D ice printing is emerging as a promising additive manufacturing method with potential applications in engineering, medicine, science, and art. Printing ice at the micrometer–millimeter scale is a challenging, high-frequency additive manufacturing process. Freeform ice 3D printing though is an early stage manufacturing process where complex geometries require significant time to design manually through an empirical trial-and-error approach, due to the nonlinear nature of droplet deposition and solidification along with unavoidable uncertainties and noise. This process would be tremendously improved with better understanding and harnessing the ability to visualize, analyze and adjust the printing process in real time. For this automation, using approaches such as closed loop control that have allowed tremendous advances in other fields like self-driving cars, plane autopilot, etc. would be very useful. In order to use closed loop control for this process, ice position and shape must be determined and transformed into in situ actionable data, which is complicated further by the transparency of ice and speed required for the phase transition from water to ice. We implemented vision techniques to build a data set to train a machine learning algorithm though an ice segmentation approach using a convolutional filter. We also implemented a hybrid optical flow algorithm (Farneback-FAST) to create a segmented video frame data set for training a neural network, Icenet. This approach was faster than the Farneback-FAST, segmenting frames in 25 ms, which allows for single and low multi droplet control. Our approaches will enable future closed loop control of the printing process and will be useful in a variety of areas including additive manufacturing, organ on a chip systems, and biomanufacturing.