基于金字塔卷积和带状池化的X光目标检测  

X-Ray Object Detection Based on Pyramid Convolution and Strip Pooling

作　　者：乔靖乾 张良 Qiao Jingqian;Zhang Liang(College of Electronic Information and Automation,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学电子信息与自动化学院,天津300300

出　　处：《激光与光电子学进展》2022年第4期209-220,共12页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61179045)。

摘　　要：安检X光图像违禁品尺度多变、姿态各异,为自动识别带来很大的困难。针对该问题,提出了一种基于金字塔卷积和带状池化的X光目标检测算法。首先,以一阶段无锚框目标检测框架CenterNet为基础,引入金字塔卷积,提出金字塔沙漏网络,丰富Hourglass-104特征提取网络的感受野,增强多尺度特征提取能力。其次,带状池化的引入能够捕捉图像上下文全局信息,防止无关区域的信息干扰,兼顾局部细节信息。最后,在训练过程中将预测目标尺度分支的训练损失替换为交并比(IoU)损失函数,进一步提升尺度预测分支的性能。消融实验结果表明,改进后网络的平均精度(mAP50)由86.6%提升为88.3%,准确率有显著提升。The scale of contraband in security X-ray image is changeable and its posture is different, which brings great difficulties to automatic identification. To address this problem, an X-ray target detection algorithm based on pyramid convolution and strip pooling is proposed. First, pyramid convolution is introduced based on CenterNet, a one-stage anchor free frame target detection framework. Then, a pyramid hourglass network is proposed to enrich the receptive field of the hourglass-104 feature extraction network and enhance the ability of multi-scale feature extraction.Second, the introduction of strip pooling can capture the global information of the image context. It can also prevent information interference of irrelevant areas and consider local detail information. Finally, to enhance the performance of the scale prediction branch in the training process, the training loss of the prediction target scale branch is replaced by the intersection over union(loU) loss function. The ablation experiment results show that the average accuracy(mAP50) of the enhanced network is improved from 86. 6% to 88. 3%, and the accuracy is significantly improved.

关 键 词：图像处理 X光图像目标检测 深度学习 金字塔卷积 带状池化 交并比损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]