InsDetecter: An Insulator Detection Algorithm Based on Foreground Focusing and Multi-Scale Interaction

In recent years, the intelligent inspection of transmission lines using drones has become increasingly important to maintain the growing total length of transmission lines. However, the complex terrain of transmission lines makes it challenging to distinguish between the foreground and background in drone imagery. Additionally, the significant scale variations in drone imagery result in variations in the features of insulators. To address these issues, we propose an improved insulator detection method based on YOLOv5. First, we introduce a Global Channel Interactive Attention (GCIA) module to establish global channel interdependencies, enhancing the semantic feature representation of insulators and enabling better differentiation from complex backgrounds. Second, we propose a Skip Feature Interaction Fusion (SFIF) module to fuse deep semantic information with shallow edge information, correcting the feature representation of insulators at different scales. In comparison to state-of-the-art models, our proposed method InsDetecter demonstrates strong accuracy and speed on the CPLID and STNPLAD datasets.

I. INTRODUCTION

With the continuous expansion of transmission line networks, there is an increasing demand for both the speed and accuracy of circuit inspections. Insulators are crucial components of transmission lines, and ensuring the quality and safety of these insulators is vital to maintaining the reliable operation of the power grid system. Currently, drone-assisted circuit inspection technology has become the mainstream method for inspections due to its efficiency and convenience. Drones can rapidly capture a large number of images of transmission line insulators. By utilizing computer-based object detection on these drone images, significant human and material resources can be saved, leading to a substantial improvement in inspection efficiency.