BEmST: Multi-frame Infrared Small-dim Target Detection using Probabilistic Estimation of Sequential Backgrounds

When infrared small-dim target images under strong background clutters are employed to train a deep learning-based detection network, the model becomes biased towards the clutters, negatively impacting detection performance. While background estimation is able to address this issue, most convolutional neural network-based ways require manual foreground mask extraction during learning phases. In unsupervised tactics, it is often assumed that the background in frames is captured by a stationary camera. Howbeit, lots of small target sequences have dynamic backgrounds due to motion in the imaging platform, challenging this hypothesis. There has limited focus on unsupervised background estimation for small target images with sensor motion. To address this gap, a learning-based model, named BEmST, is raised. BEmST combines a variational autoencoder with stable principal component pursuit optimization for unsupervised deep background modelling. Target detection is then performed using U-net++ on differences between the modelled background and input images. This innovative tactic integrates unsupervised probabilistic background estimation with supervised dense classification for bettered small target detection. Extensive qualitative / quantitative experiments on public datasets validate that BEmST not only outperforms state-of-the-art tactics in availably and robustly detecting small-dim target images across various challenging scenarios, but also achieves superior detection performance, such as higher probabilities of detection, lower false alarm rates, and larger areas under ROC curves. The results pave a way for the future utilization of small-dim target image in a more efficient manner.