基于改进YOLOv5的X射线图像危险品检测

目前安检甄别工作主要依靠人工完成，但巨大的工作量使自动安检成为必然趋势。为了在保证精度的前提下轻量化模型，并进一步克服X射线安检图像存在物品重叠遮挡、背景干扰严重等问题，提出了一种基于YOLOv5的改进模型。该模型采用Ghost模块对原模型进行剪枝处理；引入坐标注意力机制，使网络在训练中更好地聚焦危险品，生成更具分辨性的特征；采用传统数据增强策略与Mixup数据增强策略相结合的方式来改善模型的泛化能力和鲁棒性。实验结果表明：改进模型的参数量和模型大小比原模型分别减少了17.3%、16.1%,改进模型在SIXray和OPIXray数据集上的平均精度均值(mAP)比原模型分别提高了2.3%、5.7%。所提出模型在轻量化的同时能够进一步提高检测能力，满足实时检测要求，具有较好的应用价值。

Detection of Dangerous Objects in X-ray Images Based on Improved YOLOv5

At present, the security screening work is mainly done manually, but the huge workload makes automatic security inspection an inevitable trend. In order to make the model lightweight on the premise of ensuring accuracy, and further overcome the problems of overlapping objects and serious background interference in X-ray images, an improved model based on YOLOv5 is proposed. The Ghost module is used to prune the original model to reduce the model size. The coordinate attention mechanism is introduced to make the network better focus on dangerous goods during training and generate more discriminative features. A combination of traditional data augmentation strategy and Mixup data augmentation strategy is used to improve the generalization ability and robustness of the model. The experimental results show that the parameters and model size of the improved model are reduced by 17.3% and 16.1% respectively compared with the original model, and the mAP of the improved model on SIXray and OPIXray datasets is increased by 2.3% and 5.7% respectively. The proposed model can further improve the detection ability while being lightweight and meet the real-time detection requirements. It has good application value.