GDiPAYOLO: A Fault Detection Algorithm for UAV Power Inspection Scenarios
Xiao Liang, Jiali Wang, Peidong Xu, Qingyu Kong, Zhaogang Han
Abstract
Numerous real-time object detection frameworks have been evaluated using publicly available datasets, demonstrating commendable performance. However, there is a growing need for improved performance when these frameworks are applied to custom datasets. This manuscript aims to address the dual objectives of achieving high accuracy while maintaining lightweight efficiency in the context of UAV (Unmanned Aerial Vehicle) power inspection algorithm development. The newly introduced GDiBlock (Group Dimensionality-increase Convolution Block) effectively reduces channel redundancy, allowing for the extraction of rich, high-dimensional feature information with fewer parameters. Simultaneously, the PAM (Pooling Attention Mechanism) accurately distinguishes between background and target entities without introducing additional parameters, making it particularly advantageous for addressing complex background scenarios commonly encountered in UAV imagery. We meticulously explore a design paradigm guided by the receptive field principle to determine the optimal number of blocks within each processing stage. In conclusion, our proposed network architecture reduces the parameter count by 28% to 40%, while improving the [email protected] metric by a margin of 1.7 to 5.4 points when compared to YOLOv5. When compared to state-of-the-art real-time object detection frameworks, our study demonstrates superior performance in terms of both model accuracy and size