摘要
钢轨表面缺陷样本稀缺使得深度学习检测法在实际场景中表现不佳,针对传统数据增强方法得到的图像类型单一、多样性较低的问题,将注意力机制融入深度卷积生成对抗网络(DCGAN),提出一种Attention-DCGAN用于轨面缺陷图像扩充。具体地,将自我注意力和通道注意力机制融入DCGAN的生成器和判别器,自我注意力可提高图像对全局信息的利用,通道注意力增强了图像中的通道依赖关系。使用Attention-DCGAN在自制轨面缺陷数据集和Type-I RSDDs数据集上进行实验,相比DCGAN,Attention-DCGAN在自制数据集上将IS(inception score)从1.74±0.02增加到1.77±0.04,FID(Fréchet inception distance)从137.75降低到130.64;在Type-I RSDDs上将IS从1.48±0.05增加到1.54±0.02,FID从153.96降低了142.85。结果表明Attention-DCGAN在两种数据集上均提高了生成图像的质量,可用于轨面缺陷图像的扩充,有助于提高有监督深度学习检测法在钢轨检测上的应用。
The scarcity of rail surface defect samples causes poor performance of deep learning detection methodin actual applications.In response to the problems of single image type and low image diversity obtained by traditional data enhancement methods,by integrating the attention mechanism into Deep Convolutional Generative Adversarial Network(DCGAN),this paper proposed an Attention-DCGAN for rail surface defect image expansion.Specifically,the mechanisms of self-attention and channel attention were integrated into the generator and discriminator of DCGAN,with self-attention to improve the utilization of global information in images,and channel attention to enhance channel dependence in images.Experiment using Attention DCGAN was conducted on self-made rail surface defect datasets and Type-I RSDDs datasets.Compared with DCGAN,Attention-DCGAN improved inception score from 1.74±0.02 to 1.77±0.04 and decreased Fréchet inception distance(FID)from 137.75 to 130.64 on self-made datasets.On Type-I RSDDs,inception score was increased from 1.48±0.05 to 1.54±0.02,and Fréchet inception distance decreased from 153.96 to 142.85.The results show that Attention-DCGAN improves the quality of the generated images on both datasets,can be used to expand the rail surface defect image,and helps to improve the application of supervised deep learning detection method in rail detection.
作者
闵永智
李嘉峰
王果
MIN Yongzhi;LI Jiafeng;WANG Guo(School of Automation and Electrical Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;Gansu Provincial Engineering Research Center for Artificial Intelligence and Graphics&Image,Lanzhou Jiaotong University,Lanzhou 730070,China)
出处
《铁道学报》
EI
CAS
CSCD
北大核心
2023年第12期123-130,共8页
Journal of the China Railway Society
基金
国家自然科学基金(62066024)。
