Spatial-information recurrent field transform network for sequence infrared small target detection

In infrared technology, small target detection systems have garnered significant attention due to their widespread applications in fields such as search and rescue, ground target monitoring, and early warning systems. However, most existing infrared small target detection (IRSTD) methods primarily focus on single-frame images, which limits their ability to capture temporal continuity and motion cues in real-world scenarios. Single-frame approaches often struggle to distinguish true targets from background noise or dynamic interference, especially when the targets exhibit low signal-to-noise ratios, small spatial scales, or complex motion patterns. To address these challenges, we propose a Spatiotemporal Recurrent Field Transformation Network (SRFT-Net) for sequential infrared small target detection. To effectively capture and track optical flow information, we introduce a Recurrent Pixel Field Transformation module (RFT). This module extracts pixel-level features, constructs a multi-scale 4D correlation volume among all pixels, and iteratively updates the flow field (optical flow) through a recurrent unit that performs efficient lookups within the correlation volume. Additionally, we incorporate a Hybrid Attention Feature Block (HAFB), which integrates local and global attention mechanisms. This module enhances the feature extraction capability for small targets by fusing different features through attention-guided multi-path convolution, thus improving detection performance. Experimental results on the publicly available sequential infrared small target dataset (SIRSTD) demonstrate the strong detection performance of the proposed SRFT-Net. Compared to existing state-of-the-art IRSTD algorithms, SRFT-Net achieves notable improvements across multiple evaluation metrics.