基于GAN的车门上饰板表面缺陷检测数据增广算法

Data augmentation algorithm for surface defect detection on car door panels based on GAN

为了解决车门上饰板缺陷检测任务中缺陷数据集过少的问题,文章提出了一种基于生成对抗网络(generative adversarial network,GAN)的缺陷生成模型(condition defect GAN,CDGAN),主要目标是生成具有各种类型缺陷的车门上饰板图像,以增加训练数据的多样性,提高缺陷检测模型的检测性能。CDGAN模型分为两个网络:缺陷生成网络和图像转换网络。缺陷生成网络生成位于数据集边界框内的缺陷图像,图像转换网络将这些缺陷与无缺陷的图像进行合并生成缺陷图像。实验表明,生成的缺陷图像显著提高了YOLOv5模型的缺陷检测能力,在车门上饰板缺陷数据集上取得了94.9%的平均精度(mAP@50)。该方法已经应用在车门上饰板的生产制造中,在实际的工业应用中证明了该方法的可行性。

In order to solve the problem of too small defect dataset in the defect detection task of the upper door trim panel,the article proposes a defect generation model CDGAN based on generative adversarial network.The primary goal is to generate images of car door trims with various types of defects,thereby increasing the diversity of training data and enhancing the performance of the defect detection model.It involves two stages:training the GAN generator to learn the distribution of the defect dataset,and sampling data from the trained generator to augment model performance.The CDGAN model consists of two networks:a defect generation network and an image translation network.The defect generation network produces defect images located within the dataset's bounding boxes,and the image translation network merges these defects with defect-free images.Ablation experiments demonstrate that the generated defect images significantly improve the defect detection capability of the YOLOv5 model,achieving an average precision(m AP@50) of 94.9% on the car door trim defect dataset.This method has been applied in the production and manufacturing of car door trims.Its feasibility has been proven in practical industrial applications.