Accurate analysis of the spray formation process is of utmost importance because it governs the fuel injector performance in air-breathing engines. Historically, optical diagnostic measurements in this region have been limited, hindering progress in understanding the underpinning physics. Due to the rapid growth in computer power, the first principal simulations of the primary breakup process are now available, providing access to the complete spatiotemporal fields. This data is essential to investigate the physical instability modes. 

However, it typically relies on subjective criteria to identify the emerging structures in the spray formation process. Since computer vision tools are known to help automate such classification tasks, we focus on developing a framework that can be applied to help identify critical regions during the breakup process using semantic and instance segmentation techniques. The emerging spray structures identified and tracked include the well-known lobes, ligaments, and droplets that arise during fuel injection. Analysis shows that our approach yields promising results, providing helpful statistics on the fundamental breakup dynamics for engineering analysis.

Sunway University's Professor Dr Lau Sian Lun and his two research students, Lim Wei Lun and Refat Khan Pathan, have been working on this focused research for the last two years. Recently, the team presented their research outcome at the 2023 IEEE International Conference on Computing (ICOCO). These two presentations were based on outcomes from experiments conducted to detect objects from liquid spray, and object tracing was done with those detected objects.

The first research presentation was about an investigation of object detection using U-Net. The segmented result showed that semantic segmentation applied to liquid spray images led to promising outcomes. An experiment to determine how robust the trained U-Net is then carried out under different blur levels without prior knowledge. The experiment results reflect that some objects are detected better after blurring due to the smooth-out pixels and the coarse structure of the sharpened contour. 

Following this finding, a hypothesis is proposed that inducing an appropriate blurring artefact into the test image will significantly maximise the visibility of contour boundaries, allowing the U-Net to segmentise the identifiable contours efficiently. A closer inspection of the pixel histogram revealed a narrow range of high-frequency values. The reduction of pixel variation makes the object appear connected without sharp separation, improving the continuity of the contour boundary.

The second research focus presented at the conference was an instance segmentation method proposed using Mask R-CNN to detect several droplets and other object classes. After the detection, object tracing was done using a customised breadth-first search (BFS) algorithm after collecting corresponding objects using another bespoke K-nearest neighbour algorithm. 

The authors showed that their framework can successfully detect and trace the spray structures with promising accuracy and provide helpful statistics of the breakup dynamics but is limited by the object detection count. They also analyse the class transition rates and the object breakdown scenarios in different sections of the image set. The authors conclude that their framework can help understand the spray formation process and the underlying physics of atomisation. They suggest future work on improving object detection accuracy and tracking the individual droplet motion.

Professor Lau Sian Lun
School of Engineering and Technology
Email: @email

This article has been adapted from Refat Khan Pathan, Wei Lun Lim, Sian Lun Lau, Chiung Ching Ho, Luis Bravo; Rahul Babu Koneru & Prashant Khare (2024), Application of Computer Vision Techniques to Identify Spray Primary Breakup Structures in High-Speed Flow, DOI: 10.1109/ICOCO59262.2023.10398036