基于Vision Transformer的毫米波人体安检图像违禁品识别

Identification of Prohibited Items in Millimeter-Wave Human Security Inspection Images Based on the Vision Transformer

毫米波人体安检图像因为成像质量和遮挡等问题,使违禁品的识别难度较大。因此采用更优的检测识别算法,提升违禁品的识别速度和精度一直是业内重点研究的方向。将Vision Transformer(ViT)应用到毫米波图像违禁品的识别过程中,通过将无监督预训练的ViT与经典的目标检测算法(Faster R-CNN)相结合,实现了高精度的毫米波人体安检图像违禁品识别。为了充分训练和测试算法,制作一个包含枪支和刀具两类违禁品,共计14.5万个违禁品成像样本的毫米波人体安检数据集。通过与经典的基于101层残差网络(ResNet-101)的Faster R-CNN对比,该算法使mAP50提升了2.4个点,达到了89.9%。

The millimeter-wave human security inspection image is difficult to identify prohibited items because of problems such as imaging quality and occlusion,so using a better detection and identification algorithm to improve the identification speed and accuracy of prohibited items has always been a key research direction in the industry.This paper applies the Vision Transformer(ViT)to the millimeter-wave image identification process of prohibited items,combines the unsupervised pre-trained ViT with the classic object detection algorithm Faster R-CNN,and achieves the high-precision millimeter-wave human security inspection image identification of prohibited items.In order to fully train and test the algorithm,this article creates a millimeter-wave human security inspection dataset which contains two kinds of prohibited items:guns and knives,with a total of 145,000 imaging samples of prohibited items.Compared with the classic Faster R-CNN based on ResNet-101,the algorithm improves mAP50 by 2.4 points to 89.9%.