H²D-Net: High-resolution Guided Hierarchical Discriminative Network for Infrared Small Target Detection

Following detection-by-segmentation paradigm, U-net and its variants have recently achieved competitive performance in infrared small target detection (IRSTD) benchmarks. However, when the targets only occupy few pixels, the U-shape deep network tends to favor global background patterns over local appearance of targets in the feature encoding stage, and indiscriminately amplifies false feature response in the decoder. Such representation bias and error accumulation degrade identification capability when target-similar distractors occur. Here, by introducing high-resolution cues, we advocate our High-resolution Guided Hierarchical Discriminative Network (H²D-Net), where High Resolution Guidance (HRG) module and Holistic Distractor Filter (HDF) module are devised to tackle the aforementioned issues. Specifically, an extra hierarchical network with fixed scale embedding, i.e., high-resolution cues, is parallelly assigned to rectify the representation bias of the U-shape network via a group of the HRG modules, which facilitate bidirectional interaction between the fine-grained spatial details and multiscale representations. Furthermore, the refining HDF module is embedded into the bottleneck between the encoder and decoder for the purpose of interrupting feedforward propagation of the false feature response. Extensive experiments demonstrate that the H^2D-Net significantly enhances the detection performance of infrared small targets, particularly in reducing false alarms, outperforming state-of-the-art methods across multiple real-world infrared datasets.