摘要
基于深度学习的目标检测方法是当前计算机视觉领域的研究热点,但在小目标的检测问题上,基于深度学习的检测器存在较多的漏检。高光谱图像的每个像元包含了物质的光谱信息,能够提升小目标的检测率。然而,高光谱图像的相邻波段相关性高,需要从中选取具备代表性的波段以降低计算冗余。为此,提出了一种高光谱小目标检测模型,使用径向基激活函数(RBAF)进行光谱筛选与目标检测。具体而言,针对高光谱图像波段冗余的特点,利用RBAF设计注意力机制进行光谱维的特征筛选;针对小目标纹理模糊,相对于背景不显著的特点,先对输入图像进行分辨率重建,随后利用RBAF构建径向基目标输出子网络(RBOON),以加强目标分类。为了简化模型,将光谱筛选与分辨率重建整合为注意力分辨率重建子网络(ABRRN),配合径RBOON,检测模型能够筛选特定光谱,抑制虚警,从而提高小目标检测的正确率。高光谱小目标检测实验表明,本研究方法可以使两种检测精度评价指标AP50和AP50:95分别提升5.4%和0.2%,相较其他方法更具备优势。
Object detection methods based on deep learning are the current research focus of computer vision.However,when detecting small objects,existing detectors often suffer from missing detection.Every pixel of hyperspectral images contain the spectral information of small object materials.Therefore,they can provide additional support for improving the detection performance on small objects.However,the adjacent bands of hyperspectral images are highly correlated.It is thus necessary to select representative bands to reduce the computational redundancy.In response,this paper proposed a hyperspectral small object detection model,which used the radial basis activation function(RBAF)to carry out spectral screening and object detection.Specifically,in view of the band redundancy of hyperspectral images,an attention mechanism based on the RBAF was designed for spectral screening.As for the high texture fuzziness and low distinguishability against the background of small objects,the resolution of input images was reconstructed first.Then,a radial basis object output network(RBOON)based on the RBAF was constructed to enhance object classification.For model simplification,spectrum screening and resolution reconstruction were integrated into an attention-based resolution reconstruction network(ABRRN).With the combination of the ABRRN and RBOON,the detection model can screen the specific spectrum and suppress false alarms and thus improve the accuracy of small object detection.Hyperspectral small object detection experiments show that the proposed method improves the two detection accuracy criteria,namely AP50 and AP50:95,by 5.4% and 0.2%,respectively,which means it is better than other methods.
作者
王勃凡
赵海涛
Wang Bofan;Zhao Haitao(School of Information Science and Engineering,East China University of Science and Technology,Shanghai 200237,China)
出处
《光学学报》
EI
CAS
CSCD
北大核心
2021年第23期87-97,共11页
Acta Optica Sinica
基金
国家自然科学基金(61973122)。
关键词
机器视觉
图像检测系统
高光谱图像
目标检测
径向基函数
注意力机制
machine vision
image detection system
hyperspectral image
object detection
radial basis function
attention mechanism
