基于Swin Transformer和多尺度特征融合的红外弱小目标检测方法

红外弱小目标的检测识别是军事侦察和遥感探测领域的一项关键技术。针对现有的传统目标检测方法普遍存在的检测误报率高、环境适应性差等问题，本文设计提出了一种基于Swin Transformer和多尺度特征融合的红外弱小目标检测方法。该方法首先在基于编解码Unet网络架构的基础上，通过引入Swin Transformer的自注意力机制代替常规的卷积核来进行目标特征的分层提取，从而有助于在更大的感受野下挖掘目标在不同尺度下的潜在信息；之后，通过设计一个自底向上的跨层特征融合模块作为网络模型的解码器，可以从复杂背景中保留红外弱小目标特征，并将目标的浅层局部信息和深层语义信息进行充分融合。试验测试结果表明，所提方法在红外小目标公共测试数据集SIRST上能够实现0.747的交并比指标(IoU)，以及0.752的归一化交并比指标(nIoU),其性能均优于其它典型方法，在不同复杂场景下均拥有更好的检测效果。

Infrared Dim and Small Target Detection Based on Swin Transformer and Multi-Scale Feature Fusion

Infrared dim and small target detection is a key technology in the field of military reconnaissance and remote sensing. However, traditional detection methods have a high false alarm rate and poor environmental adaptability that fail to handle complex backgrounds. To address this issue, an infrared dim and small target detection method based on Swin Transformer and multi-scale feature fusion is proposed in this paper. Firstly, based on the Unet architecture, the self-attention mechanism of the Swin Transformer is introduced to replace the convolutional kernel to extract the hierarchical features of the target, which is helpful to excavate the potential information of the target in different scales under the larger receptive field. After that, through a bottom-up cross-layer feature fusion module as the decoder of the network model, the infrared small dim target features can be retained from the complex background, and the local shallow information and deep semantic information of the target can be fully integrated. The experimental results show that compared with other typical algorithms, the proposed method achieves superior performance with an intersection over union(IoU) of 0.747 and a normalized IoU(nIoU) of 0.752 on the public single-frame infrared small target (SIRST) dataset, which proves that it has better detection result in different complex scenes.