Aerial Infrared Small Target Detection via Dual-Decoder Gradient Feature Aggregation

In recent years, deep learning-based infrared small target detection (IRSTD) methods have achieved remarkable progress and become the main solution. However, complex background interference - characterized by high-intensity clutters and strong edges - often exhibits structures highly similar to infrared small targets, making precise target-background separation particularly challenging. To address the performance limitations of existing detectors, this work aims to enhance the overall detection capability by integrating a set of simple but highly effective modules within a unified architecture.Specifically, we propose a dual-decoder detection framework. This design effectively mitigates the feature resolution degradation caused by the single round of downsampling and upsampling in conventional single-decoder architectures. Meanwhile, the intermediate guidance map generated by the primary decoder provides the auxiliary decoder with more accurate target priors, thus further improving target localization. In addition, we develop an adaptive directional gradient feature extraction module, which introduces only a negligible number of parameters, but preserves rich edge and contour details in high-resolution feature maps - an ability that is particularly crucial for aerial IRSTD tasks requiring fine-grained target delineation.Extensive experiments are conducted on the public IRSTD-1K, NUAA-SIRST, and NUDT-SIRST datasets. Moreover, based on the sky-scenario subsets of these datasets, we construct a new benchmark named Air-SIRST. The experimental results demonstrate the effectiveness and superiority of the proposed method.