Haze Removal for UAV Power Line Asset Inspections using Light-weight Network
Fei Tan, Xiaoyuan Yu · 2023
Currently, unmanned aerial vehicles (UAVs) are being extensively employed for intelligent power line asset inspections. However, the visual quality of the captured video images by UAVs suffers a detrimental impact in hazy weather conditions, leading to a degradation in the accuracy of identifying equipment defects in power lines. Over the past years, researchers have proposed several methods to mitigate the adverse effects of haze on images. Nevertheless, these approaches often necessitate a considerable number of parameters and exhibit low efficiency, thereby posing challenges in meeting the real-time requirements of automatic UAV inspections. To address these challenges, this paper introduces a novel computationally efficient light-weight image dehazing network (LID-Net) specifically tailored for power-constrained scenarios, such as UAV inspections. LID-Net incorporates an encoder-decoder structure with improved depthwise separable convolutions as its basic building blocks. Moreover, it leverages a re-constructed atmospheric scattering model to jointly estimate the global information. Additionally, a dehazing benchmark specifically synthesized for power line asset inspections, named PLAI-HAZE, is proposed. Extensive experiments are conducted using hazy images, followed by quantitative and qualitative evaluations to confirm the effectiveness of the proposed method.