基于改进YOLOv8的X光图像违禁品检测

X-ray Image Contraband Detection Based on Improved YOLOv8

针对现有X光图像违禁品检测精度低的问题,提出一种基于改进YOLOv8的X光图像违禁品检测算法。在YOLOv8基础上添加了一个小目标检测头,以增强对小目标物体敏感度,同时使用自适应空间特征融合(Adaptively Spatial Feature Fusion, ASFF)模块自适应调整不同检测层间的权重,避免多尺度层级之间的信息冲突。使用重参数化后的RFB_S模块替换YOLOv8中的快速空间金字塔池化(Spatial Pyramid Pooling-Fast, SPPF)模块,使用大小不同的卷积来获取不同视野特征图信息,避免多重最大池化可能带来的梯度消失问题。在颈部网络和主干网络之间引入高效多尺度注意力(Efficient Multi-scale Attention, EMA)机制,有效区分背景区域和目标区域,加强关键信息的交互。使用可变形卷积替换C2f模块中的普通卷积,利用可变形卷积自适应调整卷积核的形状,更好地捕捉和感知图像中不同尺度的目标特征。该算法在SIXary数据集上测试平均精确度均值(mean Average Precision, mAP)达到92.7%,比原始算法提高了3.2%。实验结果表明,改进后的算法比原始算法有了较大提升,证明了改进算法的有效性。

To solve the problem of low accuracy of contraband detection in existing X-ray images,an X-ray image contraband detection algorithm based on improved YOLOv8 is proposed.Firstly,a small target detection head is added on the basis of YOLOv8 to enhance the sensitivity of small target objects.At the same time,Adaptively Spatial Feature Fusion(ASFF)module is used to adaptively adjust the weights between different detection layers to avoid information conflicts between multi-scale levels.Secondly,the Spatial Pyramid Pooling-Fast(SPPF)module in YOLOv8 is replaced with reparameterized RFB_S module,and the convolution of different sizes is used to obtain the feature map information of different fields of view,so as to avoid the potential gradient disappearance problem caused by multiple maximum pooling.The Efficient Multi-scale Attention(EMA)mechanism is then introduced between the neck network and the backbone network to effectively distinguish the background region from the target region and enhance the interaction of key information.Finally,the common convolution in C2f module is replaced by deformable convolution,and the shape of the convolution kernel is adjusted adaptively by using deformable convolution to better capture and perceive the object features of different scales in the image.The mean Average Precision(mAP)value of the proposed algorithm reaches 92.7%on the SIXary dataset,which is 3.2%higher than the original algorithm.The experimental results show that the improved algorithm is much better than the original algorithm,which proves the effectiveness of the improved algorithm.