WOODWORK: A deep-learning based framework for woodpecker damage detection in powerline inspection

Maintaining the structural integrity of wooden utility poles is essential for reliable electrical distribution and transmission systems. Woodpecker-induced damages are frequent and pose a significant hazard, compromising structural integrity and potentially jeopardizing inspection personnel’s safety and the power supply’s continuity. This paper introduces WOODWORK, a novel deep learning-based framework for segmentation and severity assessment of woodpecker damages and holes on wooden utility poles from RGB images captured by inspecting cameras. Leveraging the capabilities of the Segment-Anything Model and DINOv2, combined with Grad-CAM, our method provides a weakly-supervised approach that only requires image-level annotations, instead of pixel-level ones, to produce a model capable of specifying whether a given utility pole image contains woodpecker damages. The resulting model achieved precise localization and assessment of the severity of the damages based on the area. To further improve performance, we propose a new loss function and a novel refinement technique that significantly enhances the detection capability. Our extensive experiments demonstrate the effectiveness of WOODWORK, proving it can be a tool that can improve inspection efficiency and help reduce replacement and repair costs by facilitating early damage detection and evaluation.