融合多尺度分形注意力的红外小目标检测模型  

Infrared Small Target Detection Model with Multi-scale Fractal Attention

作　　者：谷雨[1] 张宏宇 孙仕成 GU Yu;ZHANG Hongyu;SUN Shicheng(School of Automation,Hangzhou Dianzi University,Hangzhou 310018,China;HDU-ITMO Joint Institute,Hangzhou Dianzi University,Hangzhou 310018,China)

机构地区：[1]杭州电子科技大学自动化学院,杭州310018 [2]杭州电子科技大学圣光机联合学院,杭州310018

出　　处：《电子与信息学报》2023年第8期3002-3011,共10页Journal of Electronics & Information Technology

基　　金：浙江省自然科学基金(LY21F030010);浙江省科技计划(2019C05005)。

摘　　要：为提高红外图像小目标检测的性能,融合传统方法的先验知识和深度学习方法的特征学习能力,该文设计了一种融合多尺度分形注意力的红外小目标端到端检测模型。首先,在对适用于红外图像弱小目标检测的多尺度分形特征分析基础上,给出了基于深度学习算子对其进行加速计算的过程。其次,设计卷积神经网络(CNN)学习度量得到目标显著性分布图,结合特征金字塔注意力模块和金字塔池化下采样模块,提出了一种基于多尺度分形特征的注意力模块。将其嵌入到红外目标语义分割模型时,采用非对称上下文融合机制提高浅层特征和深层特征的融合效果,并利用非对称金字塔非局部模块获取全局注意力,以提高红外小目标检测性能。最后,采用单帧红外小目标(SIRST)数据集验证提出算法的性能,所提模型交并比(IoU)和归一化交并比(nIoU)分别达到了77.4%和76.1%,优于目前已知方法的性能。同时通过迁移实验进一步验证了提出模型的有效性。由于有效地融合了传统方法和深度学习方法的优势,所提模型适用于复杂环境下的红外小目标检测。In order to improve the performance of infrared image small target detection,an end-to-end infrared small target detection model that integrates multi-scale fractal attention is designed by combining prior knowledge of traditional methods and feature learning ability of deep learning methods.Firstly,the procedure of accelerating the calculation of multi-scale fractal feature with deep learning operator is proposed based on analysis of this feature,which is suitable for detecting dim and small targets in infrared images.Secondly,the Convolutional Neural Network(CNN)is designed to obtain the target significance distribution map,and a multi-scale fractal feature attention module is proposed by combining the feature pyramid attention and pyramid pooling downsampling module.When embedding it into the infrared target semantic segmentation model,asymmetric context modulation is adopted to improve fusion performance of shallow features and deep features,and asymmetric pyramid non-local block is used to obtain global attention to improve infrared small target detection performance.Finally,the performance of the proposed algorithm is verified by experiments on the Single-frame InfRared Small Target(SIRST)dataset,where Intersection over Union(IoU)and normalized IOU(nIoU)reach 77.4%and 76.1%,respectively,which is better than the performance of the currently known methods.Meanwhile,the effectiveness of the proposed model is further verified by migration experiments.Due to the effective integration of the advantages of traditional methods and deep learning methods,the proposed model is suitable for infrared small target detection in complex environments.

关 键 词：红外小目标检测 语义分割 多尺度分形特征 注意力机制 金字塔池化下采样 

分 类 号：TN911.73[电子电信—通信与信息系统] TN219[电子电信—信息与通信工程]