A deep learning-based segmentation method for multi-scale and overlapped bubbles in gas–liquid bubbly flow

In the realm of fluid dynamics, gas–liquid bubbly flow represents a prevalent and significant multiphase flow phenomenon. With the advancement of imaging technology, high-speed photography combined with image processing techniques has become a common method for measuring bubbly flows. To overcome the challenges posed by multi-scale and overlapping bubbles in gas–liquid bubbly flows, a deep learning-based method for precise bubble contour segmentation and trajectory tracking has been developed. This approach involves specific optimizations and enhancements to the one-stage object detection model “You-Only-Look-Once version 8”, leading to a bubble segmentation algorithm that strikes a balance between speed and precision. Omni-dimension dynamic convolution and high-resolution feature layer pyramid level 2 (P2) were integrated into the model to extract more precise spatial and texture information, enhancing precision and facilitating the detection of small-sized bubbles. Additionally, to address the issue of severe bubble overlap in images, the bubble spatially enhanced attention module was developed to capitalize on detailed texture, thereby achieving the segmentation of severely overlapping bubbles. Based on the improved detection model, combined with the Botsort tracking algorithm, vanishing bubble re-identification as well as continuous tracking of severely occluded bubbles are realized. The model achieves inference speeds of 0.427 s on central processing unit and 0.03 s on graphics processing unit (GPU), respectively, facilitating its application in efficiently processing large comprehensive datasets.