【摘要】 针对复杂背景下红外小目标检测面临的目标像素占比极低、特征表示微弱、背景干扰强等挑战，以及现有深度学习方法全局建模能力不足的局限，提出一种基于频域注意力与边缘增强的检测方法。通过分层双域学习模块（hierarchical dual-domain learning module, HDLM）学习频域与空间域特征，并利用注意力机制对2类特征进行动态加权，以增强网络对小目标的定位精度与抗干扰能力；同时设计边缘增强模块（edge enhancement module, EEM），用于细化小目标边缘特征，提升网络对目标边缘的感知能力。在公共数据集IRSTD-1k上的实验结果显示，相较于当前性能领先的MSHNet（multi-scale head to the plain U-Net），所提方法的交并比（intersection over union, IoU）提升2.15百分点、检测概率（probability of detection, Pd）提升1.19百分点、虚警率（false alarm rate, Fa）降低5.88×10^-6，验证了该方法的优越性。更多还原

【Abstract】 To address the challenges of extremely low target pixel ratio, weak feature representation, and strong background interference in infrared small target detection under complex backgrounds, as well as the limitation of insufficient global modeling capability of existing deep learning methods, a detection method based on frequency-domain attention and edge enhancement was proposed. The frequency-domain and spatial-domain features were learned through a hierarchical dual-domain learning module（HDLM）, and then an attention mechanism was used to dynamically weight the two types of features to enhance the network’s localization accuracy and anti-interference ability for small targets. Meanwhile, an edge enhancement module（EEM） was designed to refine the edge features of small targets and improve the network’s ability to perceive target edges. Experimental results on the public dataset IRSTD-1k show that compared with the current state-of-the-art MSHNet（multi-scale head to the plain U-Net）, the proposed method increases the intersection over union（IoU） by 2. 15 percentage points, the probability of detection（Pd） by 1. 19 percentage points, and reduces the false alarm rate（Fa） by 5. 88×10^-6, verifying the superiority of the proposed method.更多还原