基于ScoreCAM的X光安检违禁品检测

Prohibited items detection in X-ray security inspection based on ScoreCAM

针对X光安检违禁品检测依赖数据标注,存在多姿态变化且小目标不易检出的问题,改进ResNet50网络,提出基于弱监督机制的X光安检违禁品检测模型。通过ScoreCAM算法,依靠图片类别标签实现可视化和违禁品定位,结合可变形卷积和空洞卷积,设计4种在主干网络不同位置添加可变形空洞卷积模块的网络结构,使卷积核的形态更贴近违禁品轮廓,提高违禁品特征的提取能力。仿真结果表明,该模型结果更加准确,能有效应对违禁品目标多姿态变化、遮挡、小目标漏检的技术难题。

For the X-ray security inspection of prohibited items detection relying on data annotation,there are problems including multiple postures change and small targets that are difficult to be detected.The ResNet50 network was improved and a prohibited items detection model based on weakly supervision mechanism was proposed.Based on ScoreCAM algorithm,visualization and prohibited items location was realized by image category labels.Combined with deformable convolution and dilated convolution,four network structures with deformable dilated convolution modules at different positions of the backbone network were designed.The convolution kernel shape was more similar to the contour of prohibited items and the extraction ability of prohibited items features was improved.Results of simulation experiments show that the proposed model is more accurate,which effectively solves the technical problems of multiple postures change,occlusion and small target missing detection of prohibited items targets.