摘要
针对区域建议网络中锚点框引入背景噪声导致小目标检测精度低的问题,提出了基于语义分割的感兴趣区域生成方法.首先把感兴趣区域的搜索问题转化为前景和背景的二值语义分割问题;然后对语义分割所得的前景进行中值滤波及连通域分析,直接得到感兴趣区域的大小和位置,从而避免使用锚点框来生成感兴趣区域,减小了背景噪声对目标检测的影响.对自然场景下高原鼠兔目标进行检测,结果表明:基于语义分割的感兴趣区域生成方法最优F1值比区域建议网络高27.75%,检测精度更高.
Aiming at the low precision of small target detection caused by background noise introduced by the anchor box in the regional proposal network,a method of generating interest regions was proposed based on semantic segmentation model.This method firstly transformed the search problem of the region of interest into the binary semantic segmentation problem of foreground and background.Then,the semantic segmentation directly obtained the foreground,by which the size and position of the region of interest were obtained through median filtering and connected domain analysis.This method avoided using anchor boxes to create region of interest,while reducing the impact of background noise on target detection.Experiments on the target detection of ochotona curzoniae in natural scenes show that the F1 value of the proposed method based on semantic segmentation is higher than that of the regional suggestion network by 27.75%,and the detection accuracy is higher.
作者
陈海燕
陈刚琦
CHEN Haiyan;CHEN Gangqi(Department of Computer and Com mun ication,Lanzhou University of Technology,Lanzhou 730050,China)
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2020年第7期7-12,共6页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家自然科学基金资助项目(61362034)。
关键词
高原鼠兔
语义分割
感兴趣区域
区域建议网络
无锚点框
目标检测
ochotona curzoniae
semantic segmentation
region of interest
region proposal network
no anchor box
object detection
