基于改进DCGAN轮胎缺陷图像生成方法  

A method for generating tire defect images based on improved DCGAN

作　　者：李春华 付睿智 刘玉坤[2] 王愉霖 LI Chunhua;FU Ruizhi;LIU Yukun;WANG Yulin(School of Grammar and Law,Hebei University of Science and Technology,Shijiazhuang,Hebei 050018,China;School of Information Science and Engineering,Hebei University of Science and Technology,Shijiazhuang,Hebei 050018,China;The Electrical Engineering College,Guizhou University,Guiyang,Guizhou 550025,China)

机构地区：[1]河北科技大学文法学院,河北石家庄050018 [2]河北科技大学信息科学与工程学院,河北石家庄050018 [3]贵州大学电气工程学院,贵州贵阳550025

出　　处：《河北科技大学学报》2023年第4期346-355,共10页Journal of Hebei University of Science and Technology

基　　金：河北省重点研发计划项目(21351801D);轨道交通关键装备智能运维平台研发项目(20310806D)。

摘　　要：针对深度卷积生成对抗网络的数据扩充方法存在生成图像质量差、模型框架不稳定、模型收敛速度慢等问题,提出一种改进DCGAN轮胎缺陷图像生成模型。将残差网络和注意力机制嵌入到DCGAN模型中,提升模型特征的提取能力;同时摒弃DCGAN损失函数JS散度,使用带有梯度惩罚项的Wasserstein距离,提高模型训练的稳定性。实验结果表明,使用给定模型生成的轮胎缺陷图像质量优于使用DCGAN,WGAN,CGAN与SAGAN所生成图像,其平均FID值可以达到116.28,最小FID值可以达到84.94。所提出的模型可以稳定生成质量更好的轮胎缺陷图像,为轮胎缺陷样本数据集的扩充提供了一种有效途径,有助于有效解决深度学习在缺陷检测领域发展所面临的小样本问题。An improved DCGAN tire defect image generation model was proposed to solve the problems of poor image quality,unstable model frame and slow model convergence in the data expansion method of deep convolutional generative adversarial network.The residual network and attention mechanism were embedded in DCGAN model to improve the feature extraction ability of the model.At the same time,the DCGAN loss function JS divergence was abandoned and Wasserstein distance with gradient penalty term was used to improve the stability of model training.The experimental results show that the quality of tire defect images generated by this model is better than that generated by DCGAN,WGAN,CGAN and SAGAN,with an average FID value of 116.28 and a minimum FID value of 84.94.The proposed model can stably generate tire defect images with better quality,which provides an effective way for expanding tire defect sample dataset and alleviates the problem of small sample in the development of deep learning in the field of defect detection.

关 键 词：计算机神经网络 生成对抗网络 残差网络 注意力机制 数据扩充 

分 类 号：TN958.98[电子电信—信号与信息处理]