基于增量学习的X射线安检系统检测算法研究

Detection Algorithm of X-ray Security Inspection System Based on Incremental Learning

针对现有智能X射线安检系统对新出现的异种危险物无法进行有效检测以及重新训练模型效率低的问题,研究了一种基于增量学习的X射线安检系统目标检测算法.该算法将传统目标检测网络faster rcnn中的特征提取器替换为残差网络,并在该网络的最后全连接层通过增加对应于新类的目标分类与边框回归神经元构成目标检测的增量学习网络,在该增量网络的损失函数中引入蒸馏损失解决新数据更新网络引起的灾难遗忘问题.最后,在X射线安检系统原7类数据训练模型的基础上依次增加1类新目标数据继续训练并检测,新增目标识别率不低于90%.该算法在保持网络对旧类检测能力的同时,也能将新增危险物以较高的精度检测出来.

A target detection algorithm for X-ray security inspection system based on incremental learning was studied aiming at problems that the existing intelligent X-ray security inspection system can’t effectively detect heterogeneous dangerous objects which emerges newly,and retraining is inefficient.In the method,the feature extractor of faster rcnn which was in traditional target detection network was replaced by the residual network,and an incremental learning network was constructed of Target Detection by adding target classification and border regression neurons corresponding new classes in the last fully connected layer of the network,and that distillation loss was introduced in the loss function of the incremental network to solve the catastrophic forgetting problem which was caused by updating the network with only new data.Finally,based on the original 7-class data training model of the X-ray security system,one class of new target data is sequentially added to continue training and detection,and recognition rate of the new target is not less than 90%.The experimental results show that the algorithm can detect new dangerous object with high precision while maintaining the detection ability of old classes.