Research on Height Feature Extraction Method of Tank Pool Fire Based on Improved Mask R-CNN

Abstract

Tank pool fires are a hazardous form of fire, and their height feature plays a critical role in the assessment of thermal radiation. Using Mask R-CNN as the benchmark network, this study introduces a deformable convolution structure and optimizes the loss function to achieve accurate instance segmentation of tank pool fires, extracting segmentation masks and then calculating the fire height using reference object sizes. The proposed improvements were evaluated on a self-built dataset, and the experimental results verified the effectiveness of the method. Compared to the baseline model, the improved Mask R-CNN increased AP50 and AP75 by 6.29% and 2.48%, respectively, significantly enhancing the segmentation performance of tanks and pool fires. Compared with three other segmentation algorithms, the method improved mAP by 10.95%, 7.05%, and 9.49% over Unet, Deeplab, and PSPNet, respectively, demonstrating its strong competitiveness in pool fire target segmentation.