摘要
针对包裹中的限制品在X射线安检仪中摆放角度多样且易与其他物品重叠,导致现有基于X射线影像的限制品检测与分割方法效率低、漏检及误报现象严重的问题,本文提出了基于多标签特征融合的限制品实例分割网络框架。首先,优化数据集中的弱标签,突出感兴趣区域;然后将数据集中的图像加权融合,增强网络对重叠区域的识别能力;最后,采用混合任务级联(HTC)网络实现对限制品的精确分割。在一个含有16138张图像的X射线安检数据集上的实验结果表明,本文方法相比其他方法有较高的准确率。
In an X-ray machine,the restricted items are usually placed in variety of angles and overlapped with other objects,which results in low efficiency of the existing methods based on X-ray images along with missed detection and serious false alarm problems.In view of this,the framework of network for instance segmentation on restricted items based on multi-lable feature fusion was proposed.Firstly,the weak labels in the dataset were optimized to highlight regions of interest.Then the images in the dataset were weighted and fused to increase the recognition ability of network to overlapping regions.Finally,a Hybrid Task Cascade(HTC)network was used to increase the segmentation accuracy of restricted items.Experiments are already carried out on an X-ray security dataset with 16138 images.The experimental results show that the accuracy of our method is higher than those of other methods.
作者
车翔玖
许阳
CHE Xiang-jiu;XU Yang(College of Computer Science and Technology,Jilin University,Changchun 130012,China)
出处
《吉林大学学报（工学版）》
EI
CAS
CSCD
北大核心
2020年第6期2197-2203,共7页
Journal of Jilin University:Engineering and Technology Edition
基金
国家自然科学基金项目(61672260)
吉林省科技发展计划项目(20200401077GX,20200201292JC).
关键词
计算机应用
特征融合
限制品分割
实例分割
卷积神经网络
computer application
feature fusion
restricted items segmentation
instance segmentation
convolutional neural network
