红外小目标混频特征融合检测模型

Infrared Small Target Detection with Mixed-Frequency Feature Fusion Detection Model

  • 摘要: 红外成像的小目标通常缺乏明确的轮廓和纹理信息,导致仅依靠目标自身特征进行识别存在较大困难。为克服这一不足,本文提出了一种新型混频特征融合检测(mixed-frequency feature fusion detection, MFFD)模型,它通过充分聚合目标及周边背景的上下文信息,有效提升小目标检测性能。模型中的混频提取模块通过结合全局低频语义特征与局部高频目标细节,显著增强系统对弱小目标的感知能力;此外,模型中的多阶段融合模块通过高效协同不同级别特征的交互融合,促进更深层次的语义理解和空间信息的整合。在公开数据集NUAA-SIRST和IRSTD-1k中,MFFD-Net相较于其他五种基于深度学习的方法表现更优。与AGPC-Net相比,MFFD-Net在NUAA-SIRST数据集上的IoU和nIoU指标分别提升了4.42%和4.33%,在IRSTD-1k数据集上相应指标分别提升了3.63%和6.38%。这充分表明本模型在复杂背景下进行小目标检测具有较大的应用潜力。

     

    Abstract: In infrared imaging, small targets often exhibit indistinct contours and sparse texture information, presenting a significant challenge for identification based solely on their inherent characteristics. To address this limitation, a novel mixed-frequency feature detection (MFFD) model is proposed. This model substantially improves small-object detection performance by leveraging both the contextual information of the target and its surrounding background. The MFFD model introduces a mixed-frequency extraction module that enhances small-target recognition by integrating global low-frequency semantic features with local high-frequency target details. Additionally, a multi-stage fusion module is employed to effectively coordinate feature interaction and integration across multiple levels, thereby improving semantic understanding and spatial information fusion. On the publicly available NUAA-SIRST and IRSTD-1k datasets, MFFD-Net outperformed five other deep learning-based methods. Compared to AGPC-Net, MFFD-Net achieved significant improvements in IoU and nIoU metrics. For the NUAA-SIRST dataset, increases of 4.42% and 4.33% were observed, respectively, while for the IRSTD-1k dataset, the corresponding improvements were 3.63% and 6.38%. These results demonstrate the strong potential of the proposed model for detecting small objects in complex infrared backgrounds.

     

/

返回文章
返回