Abstract: Aiming at the problems of strong background clutter, low signal-to-noise ratio (SNR) and weak target features existing in infrared dim and small target detection under complex backgrounds, a multi-scale temporal-spatial U-Net model is proposed. The model constructs a parallel spatial background modeling branch and a temporal dynamic modeling branch to model spatial structure of the background and temporal evolution characteristics of the environment respectively. An adaptive spatio-temporal fusion module is designed to dynamically fuse the features of dual branches via learnable weights. Meanwhile, target-aware weighted loss and spatio-temporal consistency constraint are introduced to effectively suppress the background while retaining the signals of dim and small targets. Experiments on multiple simulated and measured datasets show that the proposed model achieves an average detection accuracy of 97.5% and a false alarm rate as low as 2.2%, which are significantly superior to existing methods. The proposed model provides an effective solution with high precision and low false alarm for infrared small target detection in complex dynamic scenarios.

Key words: infrared small target detection, background suppression, residual network, deep learning, image recognition