摘要
现有的检测方法对轨道板细微裂缝和夜间拍摄的裂缝图像存在误检和漏检的现象,为此提出了一种基于卷积神经网络的改进方法。将特征图分组后用注意力机制强化各组向量的特征表达,以动态聚合弱分类器预测结果的方式得到最终的裂缝置信度。借助投票机制有效降低最终的预测偏差,提升模型的鲁棒性。实验结果表明:该改进方法在减少模型参数的情况下,在裂缝数据集上的准确率提升1.6%,在CIFAR-10数据集上的准确率提升2.8%。
There are misdetections and missed detections in the crack detection of track slab or in crack pictures taken at night as using existing detection methods.For the problem,an improved method based on convolutional neural network(CNN)is proposed.In this way,high-level feature maps are divided into different groups of vectors whose feature expression would be emphasized by attention mechanism subsequently.Final confidence is accounted by aggregating the predict result of weak classifiers dynamically.With the favor of voting mechanism,predict error is reduced and robustness of model is improved effectively.Experiment results show that the improved method gains a prediction improvement of 1.6%in crack dataset and an improvement of 2.8%in CTFAR-10 dataset,in spite of the reduction of model parameters.
作者
李文举
何茂贤
张耀星
陈慧玲
李培刚
LI Wenju;HE Maoxian;ZHANG Yaoxing;CHEN Huiling;LI Peigang(School of Computer Science and Information Engineering,Shanghai Institute of Technology,Shanghai 201418,China;School of Railway Transportation,Shanghai Institute of Technology,Shanghai 201418,China)
出处
《应用科学学报》
CAS
CSCD
北大核心
2021年第4期627-640,共14页
Journal of Applied Sciences
关键词
裂缝检测
卷积神经网络
投票机制
训练策略
深度学习
crack detection
convolutional neural network(CNN)
voting mechanism
training tactic
deep learning
作者简介
通信作者:李培刚,博士,讲师,研究方向为高速、重载及交通轨道结构。E-mail:lipeigang@sit.edu.cn。
