D³-Net: Derivative-Domain Discriminative Network for Infrared Small-Target Detection

Infrared imaging is pivotal for long-range aerial target surveillance; however, targets typically occupy only a few pixels and exhibit low contrast, making them easily obscured by background clutter. Existing deep learning based methods employ sophisticated architectures that aim to extract discriminative target features, yet effective extraction remains difficult due to target–background inseparability. To address this issue, we propose a derivative-domain discriminative detection framework. First, a separability-enhancing mapping is constructed to transform low-contrast pixel imagery into a derivative domain spanned by first-order slopes and second-order curvatures, which explicitly sharpens target–background structure (edge steepening and curvature-extrema focusing). On this basis, a derivative-domain discriminative network, D³-Net, is developed to model and exploit derivative-domain separable features, which includes two modules: (i)Derivative-Domain Feature (DDF) module is designed to correct mapping-induced spatial misalignment by leveraging cross-scale spatial correspondence in the derivative mapping; (ii)Derivative-Guided Attention Enhancement (DGAE) module is proposed to mitigate deep-stage attenuation of small-target cues by fusing derivative–channel priors with multi-scale spatial context. Combining all the aforementioned strategies, reliable detection of infrared small targets is expected to be achieved. Experiments on IRIS, IRSTD-1K, and SIRST-UAV show significant advantages over mainstream methods in detection accuracy, false-alarm control, and robustness to complex backgrounds under low-contrast conditions, offering a derivativespace modeling perspective for infrared small-target detection.