基于改进YOLOv8的轻量化X光图像违禁品检测  

Lightweight X-ray Image Contraband Detection Based on Improved YOLOv8

作　　者：付茂洺 赵国梁 FU Mao-ming;ZHAO Guo-liang(Civil Aviation Flight University of China,Guanghan 618000,China)

机构地区：[1]中国民用航空飞行学院,四川广汉618000

出　　处：《航空计算技术》2025年第1期22-27,共6页Aeronautical Computing Technique

基　　金：国家重点研发计划项目资助(2021YFF0603904)。

摘　　要：为了更好满足X光安检图像中违禁物品检测的实时性与可行性,提出了基于YOLOv8的轻量化检测方法,在提高检测精度的同时使模型更加轻量化。首先使用改进后的HGNetV2网络替换YOLOv8的主干网络,减少网络模型参数量、计算量和内存占用,然后使用SIoU损失函数替换CIoU损失函数,再使用注意力机制MPCA(Multi-Path Channel Attention)改进的可变形卷积DCNv2融合到颈部网络中的C2f模块,提高网络对图像关键特征的捕捉能力。其次将改进后的模型通过LAMP剪枝方案进行模型剪枝,进一步压缩模型,最后通过知识蒸馏,提升剪枝后模型的检测精度。结果显示,在公共数据集SIXary上,改进后模型相较于YOLOv8n模型mAP提高1.3%,模型参数量、计算量和模型大小分别降低66.2%、59.3%和62%。证明了改进算法的有效性。In order to better meet the real-time and feasibility of detecting prohibited items in X-ray security images,a lightweight detection method based on YOLOv8 is proposed,which improves detection accuracy while making the model more lightweight.Firstly,the improved HGNetV2 network is used to replace the backbone network of YOLOv8,reducing the number of network model parameters,computation,and memory usage.Then,the SIoU loss function is used to replace the default CIoU loss function.Finally,the attention mechanism MPCA(Multi Path Channel Attention)is used to improve the deformable convolution DCNv2 and fuse it into the C2f module of the neck network,enhancing the network's ability to capture key image features.Next,the improved model will be pruned using the LAMP pruning scheme to further compress the model.Finally,through knowledge distillation,the detection accuracy of the pruned model will be improved.The results showed that on the public dataset SIXary,the improved model showed a 1.3%increase in mAP compared to the YOLOv8n model,while reducing model parameter count,computational complexity,and model size by 66.2%,59.3%,and 62%,respectively.Proved the effectiveness of the improved algorithm.

关 键 词：目标检测 YOLOv8n 剪枝 知识蒸馏 违禁品 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]