Object tracing from synthetic fluid spray through instance segmentation

Md Refat Khan, Pathan (2024) Object tracing from synthetic fluid spray through instance segmentation. Masters thesis, Sunway University.

Preview

Text OBJECT TRACING FROM SYNTHETIC FLUID SPRAY.pdf - Accepted Version Download (9MB) | Preview

Abstract

Since the advent of rapid transportation systems, combustion engines have been a significant technological advancement. These engines require fuel to be atomized for efficient combustion. Atomization involves breaking down liquid sprays into smaller ligaments and droplets, which is crucial for optimal combustion in liquid-fueled propulsion devices. However, accurately measuring and analyzing the size and distribution of objects in a spray can be challenging, especially near the nozzle, due to the complexity of the spray and limited digital detectors. To address this, a method using two region-based R-CNN models named Mask R-CNN and BMask R-CNN and two transformer-based models named PatchDCT and FastInst has been developed to detect and segment individual droplets from synthetic liquid spray images. This approach involved training the model on augmented images and testing it on both augmented and original images. A total of 5791 epochs for Mask R-CNN, 3000 epochs for BMask R-CNN, 20k epochs for PatchDCT and FastInst have been used. Initially, the object detection rate was not good due to the dense objects all over the images. A divide and conquer technique using cropping window extraction was employed with nine window sizes to reduce object density, resulting in a significant increase in object count. The total number of objects on the first test dataset increased from 517 to 1231 using (height, width/3) and 2887 using (height/3, width/3) window with Mask R-CNN. A similar increment happened with BMask R-CNN, PatchDCT and FastInst model testing. After filtering all of the crop windows, it is shown that the models are able to identify significantly more droplets while maintaining a good mask prediction when the width is decreased by three while maintaining the height as the fluid flow is horizontal. The resulting objects were then analyzed using a customized nearest neighbor algorithm to calculate their correspondence between frames and a BFS algorithm was used to trace their movement path. On the training dataset, Mask R-CNN achieved 74.93 mean average precision with 75% IoU, where BMask R-CNN has only 44.68, PatchDCT shows 81 and FastInst achieved 41.49. Along with the movement path, objects’ contact tracing and breakdown rate has also been calculated. The information obtained, including object size, distribution, and tracing, can be valuable for engineers designing fuel injectors, leading to improved performance and efficiency.

Actions (login required)

View Item