摘要
零样本目标检测是近年来用于对训练中未见(unseen)类别目标进行分类和定位的一种技术。由此带来了目标检测中的新问题——目标视觉特征与其对应的类别语义信息映射关系不准确,未见类别目标与背景分辨性不强。提出的强分辨算法(adaptive channel with center distance Gaussian distribution loss,ACDG)使用特征图通道自适应加权机制,能够根据不同类别目标对特征图通道自动赋予权值,加强关键特征图的权重,抑制次要特征图的信息,增强特征提取网络的表征性,以建立更准确的视觉特征和语义特征之间映射关系。为了解决未见类别目标与背景分辨性不高的问题,提出中心距离高斯分布损失,约束预测边界框中心点与真实值(groundtruth)中心点位置的距离,进而加快损失函数收敛。为了验证所提算法的先进性,在MS COCO数据集上完成了大量实验,召回率和平均精度分别高出原始方法5.9%和4.5%。
There are two main problems in zero-shot object detection—the correspondence between the visual features of the object and its corresponding category semantic information is not accurate,and the discrimination between the unseen category object and the background is confused.The proposed ACDG algorithm adopted the adaptive weighting mechanism of feature map channels,which could automatically assign weights to feature map channels according to significance,to establish more accurate visual-semantic correspondence relations.In order to solve the problem of low discrimination between the unseen objects and the background,the method introduced the center distance Gaussian distribution loss to constrain the distance between the center point of the boundary frame and the groundtruth,thus speeding up the convergence of the loss function.In order to verify the superiority of the proposed algorithm,a large number of experiments were performed on MS COCO dataset,and the recall rate and average accuracy are 5.9% and 4.5% higher than the baseline,respectively.The proposed method has positive significance for improving the performance of zero-shot object detection.
作者
宋雨
李敏
何玉杰
苟瑶
吕奕龙
贺翥祯
Song Yu;Li Min;He Yujie;Gou Yao;Lyu Yilong;He Zhuzhen(College of Information&Communication,National University of Defense Technology,Wuhan 430030,China;College of Operational Support,Rocket Force University of Engineering,Xi’an 710025,China)
出处
《计算机应用研究》
CSCD
北大核心
2022年第11期3475-3480,3515,共7页
Application Research of Computers
基金
国家自然科学基金资助项目。
关键词
零样本目标检测
通道自适应
高斯分布
zero-shot object detection
adaptive channel
Gaussian distribution
