Real-Time Infrared Horizon Detection in Maritime and Land Environments Based on Hyper-Laplace Filter and Convolutional Neural Network

The infrared (IR) horizon is an essential reference for near-ground state awareness and target detection. The traditional horizon detection algorithms based on hand-crafted features claim the advantage of being intuitive and highly targeted, however, their capability of detail extraction and robustness is not strong enough when switching among diverse scenes, meanwhile, the convolutional neural networks (CNNs) have the powerful ability in feature acquisition, but has the disadvantages of interpretability and instability. Given this, this article proposes a novel horizon detection algorithm that combines the traditional method and CNN. First, the hyper-Laplace filter (HLF) is proposed to eliminate the interferences and enhance the saliency of the target. Then a positioning module (PM) based on the traditional idea which fully utilizes the structural characteristics of multifeature-maps in CNN is present to extract the endpoints of the horizon. Finally, an auxiliary module that consists of a revise branch (RB) and self-attention module (SAM) is designed to assist the PM with detailed corrections. Compared with the state-of-the-art algorithms, experiments based on three datasets with more than 25000 frames under various scenes demonstrate that the proposed algorithm not only achieves the best accuracy and highest stability but also addresses real-time needs. Notably, the detection deviation is controlled within 2 pixels on all three datasets while achieving a comparable detection speed of 170 frames/s. Besides, our algorithm also shows strong robustness even under harsh sea conditions and urban backgrounds. Code and trained models are available at https://github.com/FJsRepo/InfML-HDD.