【摘要】 针对表面缺陷检测在实际工业场景中的诸多应用需求，提出一种基于多任务深度学习的表面缺陷检测方法。不同于常规的单任务缺陷检测方式，该方法通过共享语义分割任务与分类任务的特征提取结构，实现同时处理两种任务，借助分割任务加强模型的特征提取能力，充分利用现有标签，提高分类任务检测效果，并对模型的训练策略进行优化，平衡多任务推理需求。在实际工业现场采集的图像数据集上进行了测试，结果表明与常规分类网络和目标检测网络相比，该方法对不同类型的表面缺陷均具有较好的检测效果，且模型的检测效率更加高效、训练策略更加合理有效。更多还原

【Abstract】 To meet the diverse application requirement needs for surface defect detection in real industrial scenarios, a surface defect detection method based on multi-task deep learning is proposed. Different from the conventional single task defect detection model, the proposed method enables simultaneous processing of both semantic segmentation and classification tasks by sharing their feature extraction backbone. By leveraging the segmentation task to enhance the model’s feature extraction capability, the proposed method fully utilizes existing labels to improve classification performance while optimizing the training strategy to balance inference requirements across multiple tasks. The images collected from the actual industrial field underwent testing and analysis, showing that proposed method demonstrates superior detection capabilities for various types of surface defects compared to the conventional classification network and target detection network. Furthermore, the model achieves higher detection efficiency and a more rational and effective training strategy.更多还原