Foreground-Aware Region of Interest Cropping and CBAM Attention for Excavator Segmentation in UAV Imagery

Abstract

In disaster situations like earthquakes, excavators have a crucial part to play in the rapid clearing of rubble, which is critical for conducting search and rescue efforts. Effective tracking and localization of excavators can be achieved through accurate monitoring of their presence and distribution in disaster-hit regions, enabling better coordination and resource management. This study presents a deep learning-based approach for the segmentation of excavators in high-resolution Unmanned Aerial Vehicle (UAV) imagery, addressing challenges related to small object representation and complex backgrounds. Specifically, our proposed approach involves using a foreground-aware region of interest cropping mechanism to ensure that there are sufficient excavator pixels in the training dataset, and a convolutional block attention module to improve feature refinement by taking into account channel-wise and spatial relationships between features. The proposed network utilizes a U-Net architecture based on ResNet-34. Experimental evaluations on UAV imagery reveal that the proposed approach achieved an IoU of 0.663, mIoU of 0.830, Dice score of 0.798, Precision of 0.856, Recall of 0.747 and geometric mean (G-Mean) of 0.863. These results demonstrate a notable improvement over the baseline model (U-Net with a ResNet-34 backbone), with the IoU increasing from 0.509 to 0.663 and the Dice score improving from 0.675 to 0.798.