AER-Net: Adaptive Feature Enhancement and Hierarchical Refinement Network for Infrared Small Target Detection

The infrared small target detection (IRSTD) has significant value in many civilian and military applications. However, most current methods still struggle to maintain tiny targets with low contrast and are prone to false alarms caused by confusing noise similar to the targets. Furthermore, challenges remain in accurately segmenting target shapes due to the difficulty in acquiring sufficient fine-grained details. To address these issues, we propose the adaptive feature enhancement and hierarchical refinement network (AER-Net) to extract discriminative features of small targets while leveraging the distinct semantics of different layers for reducing false alarms and preserving edge contours of targets. Specifically, we present a U-shaped architecture tailored for IRSTD by a feature representation enhancement module (FREM), which combines the local correlation of convolution neural networks (CNNs) and the global correlation of Transformers to enhance features while preventing the loss of tiny targets. Based on this, we further devise a daisy-shaped correlation-aware module (DCM) to capture sufficient global contexts for mitigating clutter interference in deeper layers. Subsequently, a gated refinement module (GRM) is designed to refine low-level features from the encoder utilizing high-level features as guidance for preserving local details and further removing noise in shallower layers. Moreover, we integrate our method with the monitoring device to extend the application of IRSTD to automatic monitoring of birds in plateau lakes and wetlands, while contributing an infrared bird dataset named BSIRST_v1 with targets exhibiting high morphological uncertainty to advance research in this field. Extensive experiments on three public and the proposed datasets have demonstrated that our method achieves a balance between detection capability and segmentation accuracy. Notably, our approach has significant improvements in the mean intersection over union (IoU) metric on the challenging benchmark NUDT-SIRST and IRSTD-1k datasets with 4.46% and 3.86%, respectively. The source code and dataset are available at: https://github.com/fuqingzhang/AER-Net .