Explore Hybrid Modeling for Moving Infrared Small Target Detection
Mingjin Zhang,Shilong Liu,Yuanjun Ouyang,Jie Guo,Zhihong Tang,Yunsong Li
DOI: https://doi.org/10.1145/3664647.3680887
2024-01-01
Abstract:Moving infrared small target detection, crucial in contexts like traffic management and maritime rescue, encounters challenges from factors such as complex backgrounds, target occlusion, camera shake, and motion blur. Existing algorithms fall short in comprehensively addressing these issues by exploring hybrid modeling, impeding generalization in complex and dynamic motion scenes. In this paper, we propose a hybrid modeling method for moving infrared small target detection via smoothed-particle hydrodynamics (SPH) and Markov decision processes (MDP). SPH can simulate the motion trajectories of targets and background scenes, while MDP can optimize detection system strategies for optimal action selection based on contexts and target states. Specifically, we develop an SPH-inspired image-level enhancement algorithm which models the image sequence of infrared video as a 3D spatiotemporal graph in SPH. Enhancing the spatiotemporal information of the target using the designed sliding window and fluid dynamics formula. In addition, we design an MDP-guided temporal feature perception module. This module selects reference frames, aggregates features from both reference frames and the current frame. The previous and current frames are modeled as an MDP tailored for multi-frame infrared small target detection tasks, aiding in detecting the current frame. Conducted extensive experiments on two public dataset: DAUB and DATR, the proposed network surpasses the state-of-the-art methods in terms of objective metrics and visual quality.