FANM: Fuzzy-Aware Nested Mamba for Infrared Dense Small Target Detection

The task of infrared dense small target detection aims to accurately locate densely distributed thermal radiation targets in complex scenes. However, in complex scenes, dense small targets often suffer from occlusion, shadows, and blurriness, leading to unclear detection bounding boxes and inaccurate localization. To address the aforementioned issues, this paper introduces an innovative approach for infrared dense small target detection, termed the Fuzzy-Aware Nested Mamba (FANM) network. Specifically, the network takes a single-frame infrared image as its input and leverages shallow, middle, and deep Mamba networks to fully exploit and integrate local and global multi-scale features. The extracted features are subsequently passed to a feature pyramid network (FPN) to balance the spatial resolution and semantic information of each feature map layer, ensuring that feature maps of different scales can be effectively utilized. Then, the output features at each level are fed into the detection head, which consists of two branches: one for regressing the positions of arbitrarily distributed bounding boxes, and the other for predicting the quality of the bounding boxes. The proposed fuzzy perceptual loss function is used to optimize these branches, ultimately outputting images with accurately detected target bounding boxes. Experiments were performed on the DMIST dataset, which includes images with varying numbers of targets, and the results validate the effectiveness of our method.