基于CSA-StyleGAN的连铸坯表面缺陷图像生成算法  

Surface Defect Image Generation Method of Continuous Casting Billet Based on CSA-StyleGAN

作　　者：杨昆[1] 徐锟 吴守仓[2] 张之江[1] YANG Kun;XU Kun;WU Shoucang;ZHANG Zhijiang(School of Communication and Information Engineering,Shanghai University,Shanghai 200444,China;MCC Baosteel Technical Services Co.,Ltd.,Shanghai 200941,China)

机构地区：[1]上海大学通信与信息工程学院,上海200444 [2]中冶宝钢技术服务有限公司,上海200941

出　　处：《电视技术》2024年第8期13-19,共7页Video Engineering

摘　　要：使用计算机视觉和深度学习技术对生产的连铸坯表面缺陷进行自动化检测,需要基于大量数据训练模型,而现场缺陷数据的获取较为困难。针对连铸坯表面缺陷数据短缺问题,提出一种基于CSA-StyleGAN的连铸坯表面缺陷图像生成算法。该算法基于StyleGAN单阶段训练,在判别器中引入注意力机制。多类别样本的联合训练过程中,使用多样性损失加权计算生成器损失函数,增强模型对各样本类别的关注度。实验结果表明,与现有的生成式对抗网络相比,所提算法的弗雷切特初始距离(Frechet Inception Distance,FID)指标平均降低60%左右,且模型训练成本平均下降50%。When using computer vision and deep learning technology to automatically detect surface defects in continuously cast slabs,it is necessary to train the model based on a large amount of data,and it is difficult to obtain on-site defect data.In response to the shortage of continuous casting billet surface defect data,this paper proposes a continuous casting billet surface defect image generation algorithm based on CSA-StyleGAN.This algorithm is based on StyleGAN single-stage training and introduces an attention mechanism in the discriminator.During the joint training process of multi-category samples,diversity loss weighting is used to calculate the generator loss function to enhance the model's attention to each sample category.Experimental results show that compared with the existing generative adversarial network,the Frechet Inception Distance(FID)index of this algorithm is reduced by about 60%on average,and the model training cost is reduced by 50%on average.

关 键 词：缺陷检测 深度学习 图像生成 注意力机制 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论]