结合异常检测的X射线安检图像识别方法

针对Faster R-CNN模型对X射线安检图像中危险品检测准确率低、误检率高的问题,提出了一种前置预分类头部的X射线安检图像检测网络(pre-classification Faster R-CNN,PC-Faster R-CNN)。该模型在骨干网络之后,区域建议网络(Region Proposal Networks, RPN)之前新增一个预分类模块先对X射线安检图像进行异常检测,提高模型对正常图像的识别能力;同时引入兴趣区域对齐(RoIAlign)模块,减小兴趣区域池化层(RoIPooling)引起的量化误差,进而提升Faster R-CNN的检测性能。新模型将浅层卷积层的低级边缘特征输入到预分类模块,使其学习正常图像的高级语义特征,从而改善整个模型的识别性能。实验结果表明,与原始的Faster R-CNN相比,本文模型对危险品的检测精度提升了9.03%,误检率降低了24.03%;同时预分类头部使得模型较大地提高了检测效率,比原始的Faster R-CNN提升了44.54%。

X-ray security image recognition methodcombined with anomaly detection

Aiming at the problem of Faster R-CNN''s low detection accuracy and high false positive rate of dangerous goods in X-ray security images, a kind of X-ray security image detection network with a pre-classification head (Pre-classification Faster R- CNN, PC-Faster R-CNN) was proposed in this paper. A new pre-classification module that shares the backbone for anomaly detection was added to improve the model''s ability to recognize normal images. RoIAlign is also introduced to reduce the quantization error caused by RoIPooling, which in turn improves the detection performance of Faster R-CNN. The features from low-level convolutional layer were fed to the pre-classification module, allowing it to learn high-level semantic features of normal images and improve the recognition performance. The results show that compared with the original Faster R-CNN, the mean average precision is improved by 9.03%, the false positive rate is reduced by 24.03%. With the pre-classification module, the detection efficiency was also improved by 44.54% over the original Faster R-CNN.