逐像素注意力驱动的红外小目标检测网络

Pixel-wise attention driven infrared small target detection network

红外小目标检测在军事和民用领域获得了广泛应用,但其存在目标尺度小、细节少、复杂背景干扰等问题,现有经典深度学习检测方法往往适用于通用目标检测,对红外小目标适配性较差。针对上述问题,构建了一种新的基于U形注意力块和逐像素注意力块的红外小目标检测网络。设计了U形注意力块,在单层级内通过局部U形子网络提取多尺度特征,并通过逐像素注意力精细化增强小目标特征,丰富多尺度小尺度目标特征表示,提升网络对小尺度目标判别能力;通过稠密融合方式进一步保留小目标信息,缓解不同层特征融合时的语义鸿沟,降低漏检率;将空间与通道2个维度逐像素注意力块应用于融合后的特征图,避免小目标特征被衰减,同时抑制复杂背景干扰。实验结果表明,提出的网络在2个红外小目标数据集NUDT-SIRST与IRSTD-1k上交并比、检测概率、虚警率指标均超过最新基准方法。此外,所提网络在检测精度和效率上也达到较好平衡。

Infrared small target detection has been widely used in the military and civil fields.The detection difficulties lie in small target size,few details and complex background interference.Existing classic deep learning detection methods are commonly suitable for generic target detection,but have poor adaptability to infrared small targets.To handle these problems,this article build a new infrared small target detection network based on U-shaped attention block and pixel-wise attention block.Firstly,a residual U-shaped attention block to extract multi-scale features through local U-shaped subnetworks in a single layer level and extract multiscale features to enrich the representation of small-scale target features is designed,so as to enhance the network′s ability to discriminate small-scale targets.Then,small target information is further preserved through dense fusion method to alleviate the semantic gap feature fusion between different layers and reduce the miss detection rate.Finally,the pixel-wise attention in space and channel dimensions is applied to the fused feature map to enhance small targets and suppress complex background interference.The experimental results show that the present network outperforms the latest benchmark method in the intersection over union,detection probability and false alarm rate of two infrared small target data sets NUDT-SIRST and IRSTD-1k.Moreover,the present network also achieves a good balance between the detection accuracy and the efficiency.