Adaptive Adjustment Multiteacher Distillation Network for Visible–Infrared Transmission Line Detection

Transmission line detection (TLD) is crucial for maintaining a stable power supply across different regions. However, manual detection of transmission lines in complex terrains poses challenges. With the rapid development of computer vision, TLD using drones has been proposed to replace manual transmission line inspection. One potential approach involves using a two-stream network that incorporates auxiliary images, such as infrared (IR) or depth images. Although this approach has demonstrated promising results, its ability to detect small objects needs improvement and the model needs to be simplified. This study introduces an adaptive adjustment multiteacher distillation network called AMDNet-S∗. It is designed for visible image (VI)-IR TLD∗. A two-stream-based AMDNet-T1 and a single-stream-based AMDNet-T2are proposed as teacher networks. They use the Shunted-B backbone to guide the learning of the lightweight single-stream RGB student network AMDNet-S∗ [without knowledge distillation (KD)], which is deployed using the Shunted-T backbone. This network obtains precise dimensional information efficiently while ensuring excellent detection outcomes on nonthermal imaging devices. Extensive TLD experiments demonstrate the superiority of AMDNet-T1∗ and AMDNet-T2∗. AMDNet-S, an enhancement of the student network using our distillation technique, consists of only 11.9 M parameters, which is a significant reduction compared to the 79.62 and 40.83 M parameters of the two teacher networks. This reduction leads to significant improvements in computational speed.