融合双向路由注意力的多尺度X光违禁品检测  

Multiscale X-Ray Contraband Detection Incorporating Bidirectional Routing Attention

作　　者：王若璇 李野[1] 赵鹏[1] 

机构地区：[1]长春理工大学物理学院,吉林 长春

出　　处：《计算机科学与应用》2024年第3期78-95,共18页Computer Science and Application

摘　　要：针对违禁品检测中存在的复杂背景干扰、物体间的重叠遮挡和多尺度变化问题,提出一种基于改进YOLOv7的X射线违禁品目标检测算法。首先,在主干中引入MBConv,以更有效的捕获全局信息;其次在特征融合网络中加入RFE模块,以增加特征图的感受野,从而提高违禁品多尺度检测的准确性。并设计出一种ELAN-BiF模块,用于抑制复杂背景干扰,使网络提取不同尺度的物品特征;为了提高小目标物体的检测精度,增加了一个微小物体检测头;最后,结合CARAFE上采样和Mish激活函数来提高网络对重叠和遮挡对象的识别能力,并提升在正负样本不平衡情况下的检测能力。结果表明,改进后的模型在SIXray_OOD数据集上进行测试,该方法map达到了95.2%,比原模型提高4.9%,比其他主流检测模型在违禁品检测任务上具有更好的优越性。Aiming at the problems of complex background interference, overlapping occlusion between ob-jects and multi-scale change in contraband detection, an X-ray contraband target detection algo-rithm based on improved YOLOv7 was proposed. Firstly, MBConv is introduced into the backbone to capture the global information more efficiently;secondly, an RFE module is added into the feature fusion network to increase the receptive field of the feature map, so as to improve the accuracy of contraband multi-scale detection. And an ELAN-BiF module is designed to suppress the complex background interference so that the network ex-tracts the features of items at different scales;In order to improve the detection accuracy of small target objects, a small object detection head has been added;finally, CARAFE up-sampling and Mish activation function are combined to improve the network’s ability to recognize overlapping and occluded objects and enhance the detection ability in the case of positive and negative sample imbalance situation. The results show that the improved model is tested on the SIXray_OOD dataset, and the method achieves a map of 95.2%, which is 4.9% better than the original model, and has a better superiority than other mainstream detection models in the contra-band detection task.

关 键 词：X射线图像 双向路由注意力 小目标检测层 YOLOv7 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]