电子测量与仪器学报

邢琛聪,吕淑静,吕岳.基于小样本学习的X光图像违禁物品识别[J].电子测量与仪器学报,2021,35(3):204-210

基于小样本学习的X光图像违禁物品识别

Few shot method for prohibited item inspection in X ray images

DOI：

英文关键词:prohibited item inspection X ray images image segmentation few shot learning metric learning

基金项目:上海市科委(18DZ2270800)项目资助

作者	单位
邢琛聪	1.华东师范大学计算机科学与技术系上海200062；
吕淑静	2.华东师范大学上海市多维度信息处理重点实验室上海200241
吕岳	2.华东师范大学上海市多维度信息处理重点实验室上海200241

Author	Institution
Xing Chencong	1. School of Computer Science and Technology, East China Normal University, Shanghai 200062, China；
Lv Shujing	2. Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University, Shanghai 200241,China
Lv Yue	2. Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University, Shanghai 200241,China

摘要点击次数: 1374

全文下载次数: 1435

中文摘要:

自动X光安检是维护公共安全的重要手段。现有X光违禁物品识别的研究仅针对数据集包含的类别，不能直接应用于未参与训练的类别。数据集各类别的数量不平衡也会影响模型的性能。针对以上问题，提出一种基于小样本学习的X光图像违禁物品分割方法。模型首先将测试图像与标注的支持图像映射至相同的特征空间，然后度量图像间的像素级相似度与区域级相似度，最后根据特征相似度分割测试图像内违禁物品区域。针对不定数量的支持集图片，采取基于卷积化门控循环单元的操作，将测试图像与不同支持图像的相似度信息融合。实验结果表明，模型在单张标注图像支持集（1 shot）和5张标注图像支持集（5 shot）情况下的准确率相比现有最优方法分别提高20%和22%，进一步证明模型具有扩展至新类别的能力。

英文摘要:

Automatic X ray security inspection is an important method to maintain public safety. Current research of prohibited item inspection on X ray images only works on pre defined classes in the dataset and cannot be generalized to unseen categories. The imbalance problem in the dataset will also affect the performance of models. In order to solve above defects, the paper proposes a segmentation model for prohibited item inspection in X ray Images based on few shot learning. The model first embeds the test image and annotated support images to a common space, then measures the spatial pixel wise similarity and regional similarity, finally segments out suspected areas in the test image. To deal with uncertain numbers of support images, a fusion method based on the ConvGRU is proposed to integrate the similarity information for the test image and different support images. Experiments show that the proposed model improves 20% and 22% meanIoU compared to the state of the art methods under 1 shot task and 5 shot task, which demonstrates the ability to recognize unseen categories.

查看全文查看/发表评论下载PDF阅读器