ACANet: attention-based context-aware network for infrared small target detection

Small infrared targets in complex backgrounds are easily obscured by a large amount of clutter, so highly discriminative features are needed to distinguish them from the cluttered background. Even if the expansion factor is increased, methods based on semantic segmentation are still difficult to aggregate local features of obscure and small objects. Even by increasing the dilation factor, methods based on semantic segmentation still have difficulty in aggregating local features of obscure and small objects. In order to effectively separate targets from their background environments and extract information of small targets with complex backgrounds, this paper proposes the attention-based context-aware network (ACANet) model. Context-aware feature aggregation model adopts dilated convolutions to extract high-level features, yet mitigates potential harm to small targets by clustering local features and reducing the dilation factor. Subsequently, abstract and detailed information are effectively fused through a cross-layer feature fusion model based on cross-layer attention and asymmetric feature fusion methods. In addition, the global pixel perception module is also used to assist in capturing target position information, which can alleviate the precise positioning challenge of the model due to the small size of the target. Experimental results show that compared with other state-of-the-art methods on existing public datasets, ACANet exhibits superior performance in metrics such as detection probability, false alarm rate, and mIoU.