胡文杰,吴晓波,李波,徐天伦,姚为.基于Self-Attention的单样本ConSinGAN模型的工业缺陷样本图像生成[J].中南民族大学学报自然科学版,2022,41(3):356-364
基于Self-Attention的单样本ConSinGAN模型的工业缺陷样本图像生成
Single sample image generation of industrial defect samples based on self-attention ConSinGAN
  
DOI:10.12130/znmdzk.20220315
中文关键词: 单样本  图像生成  注意力机制  工业缺陷  对抗式生成神经网络
英文关键词: single sample image  image generation  attention mechanism  industrial defects  GAN
基金项目:国家自然科学基金资助项目( 61976226) ; 中南民族大学科研学术团队资助项目( KTZ20050)
作者单位
胡文杰 中南民族大学 计算机科学学院武汉 430074 
吴晓波 中国外运股份有限公司北京 100029 
李波 中南民族大学 计算机科学学院武汉 430074 
徐天伦 中南民族大学 计算机科学学院武汉 430074 
姚为 中南民族大学 计算机科学学院武汉 430074 
摘要点击次数: 14
全文下载次数: 14
中文摘要:
      在工业生产中,基于机器视觉的产品质量检测方法已逐步引入生产线,但绝大多数检测模型都需要充足的缺陷样本集以完成训练.随着生产工艺的改进,缺陷样本出现的概率逐渐降低.缺陷样本过少导致工业缺陷的检测或分割任务难以实施模型训练.采用GAN模型进行样本生成可以有效增广训练样本集.主要研究了面向小样本集的工业缺陷样本生成模型,对单张样本生成网络ConSinGAN进行了改进,引入双通道自注意力机制,增强了单张样本对缺陷区域的学习能力;并引入结构相似度量改进损失函数,提高了工业生成样本中背景的纹理一致性.实验结果表明:所提出的改进后生成式对抗网络模型在单样本工业缺陷图像生成上更为稳定有效.
英文摘要:
      In industrial production, product quality inspection methods based on machine vision have been gradually introduced into the production line, but most inspection models require sufficient defect sample sets to complete training. With the improvement of the production process, the probability of defective samples gradually decreases. Few defect samples make it difficult to implement model training for industrial defect detection or segmentation tasks. Using GAN model for sample generation can effectively increase the training sample set. The industrial defect sample generation model for small sample sets,is studied, the single sample generation network ConSinGANand introduced a dual-channel self-attention mechanism to enhance this learning ability of a single sample on the defect area is improved. Applying the structural similarity measure is effective to improve the loss function which improves the texture consistency of the background in the industrial generated samples. The experimental results show that the improved generative confrontation network model proposed is more stable and effective in the generation of single-sample industrial defect images.
查看全文   查看/发表评论  下载PDF阅读器
关闭