基于改进SSD的X光图像管制刀具检测与识别

以安检X光图像管制刀具自动检测识别系统为研究对象,针对原始SSD(Single Shot MultiBox Detector)算法对浅层特征图表征能力不强,在训练阶段小目标特征逐渐消失,检测精度与实时性不佳,存在对安检危险品中管制刀具等小目标漏检误检等问题,从两个方面对原始SSD进行改进:一方面,用抗退化性能更强的ResNet34网络替换SSD中的基础网络VGG16,构建SSD-ResNet34网络模型,对基础网络后三层作卷积并进行轻量级网络融合,形成新的低层特征图;将网络部分扩展层作反卷积,形成新的高层特征图。另一方面,采用跳跃连接的方式将高层特征图和低层特征图进行多尺度特征融合。经实验分析,改进后的算法对X光图像管制刀具等小目标的检测精度和速度均有明显提升,且算法鲁棒性好,实时性良好。在VOC2007+2012通用数据集上,改进SSD算法的检测精度比SSD算法高1.7%,达到了80.5%。

X-Ray Image Controlled Knife Detection and Recognition Based on Improved SSD

In the automatic X-ray imaging systems used to identify knives in security check,using the original single shot multibox detector(SSD)algorithm,the shallow feature maps are poorly represented,features of small targets gradually disappear during the training stage,leading to low detection accuracy and poor real-time performance,and the small targets such as the controlled knives in security check are missing and checked out by mistake.To solve this problem,the original SSD was improved in two ways.On the one hand,the SSD-Resnet34 network model was constructed by replacing the basic network VGG16 in the SSD using a ResNet34 network with stronger anti-degradation performance,and the last three layers of the basic network were convolved and a new low-level feature map was created by lightweight network fusion.Part of the extended layer of the network was deconvolved to form a new high-level feature map.On the other hand,jumping connection was adopted to achieve multi-scale feature fusion between the high-level feature map and the low-level feature map.Analysis of test data shows that the improved algorithm demonstrates improved detection speed and detection accuracy of small targets,such as the X-ray image controlled knives.And the algorithm demonstrates improved robustness and high real-time performance.Using the VOC2007+2012 general dataset,the detection accuracy of the improved SSD algorithm is 1.7%higher than that of the SSD algorithm,reaching 80.5%.