Treffer: 伪标签校正和优化的跨场景带钢缺陷识别.

The data distributions of cross-scenario images of strip steel defects vary considerably due to imaging factors such as camera type, parameters, and environmental illumination, resulting in poor generalization performance of defect recognition models based on deep learning. To address this issue, we propose a pseudo label correction and optimization domain adaptation ( PLCODA) model for strip steel defect recognition. Firstly, a Retinex image enhancement module based on maximum entropy and brightness constraints was designed to generate an intermediate domain that is consistent with the label information in the source domain while different from the data distribution in the two domains. Second, we develop a dual-prediction adversarial coupling architecture that performs adversarial learning between the target domain and each of the source and intermediate domains to generate initial pseudo-labels for target-domain samples. Finally, we propose a pseudo-label correction and iterative purification strategy: we correct pseudo-labels via an improved noise matrix, then iteratively purify them by reinforcing high-confidence predictions, self-punishing low-confidence predictions, and reducing the discrepancy between pseudo-labels and ground-truth labels using a designed label-difference metric. The method was validated on steel-strip defect datasets from Handan Iron & Steel Group and the publicly Severstal Steel Defect Dataset. Experimental results show that the proposed method is superior to the existing domain adaptation methods for cross-scenario defect recognition. [ABSTRACT FROM AUTHOR]

由于摄像机类型、参数和环境光照等因素影响, 在跨场景情况下采集的带钢图像数据分布存在较大差异, 基于深度学 习的缺陷识别模型泛化性能差。 为此, 提出一种伪标签校正和优化的带钢缺陷识别域适应模型 (pseudo label correction and optimization domain adaptation, PLCODA)。 首先, 设计一种基于最大熵和亮度约束的 Retinex 图像增强模块, 用于生成一个与源域 标签信息一致且不同于两域数据分布的中间域。 其次, 设计一个双预测对抗耦合模型, 通过源域和中间域分别与目标域的对 抗学习, 实现对目标域样本的初始伪标签生成。 最后, 设计一种伪标签校正与迭代提纯策略, 通过改进噪声矩阵对伪标签校 正, 并利用高置信度预测的迭代提纯、低置信度预测的自我惩罚和设计的标签差异度量函数减小伪标签与真值标签的差距。 在邯郸钢铁集团和公开的谢维尔钢铁集团的带钢缺陷数据集上进行了验证, 实验结果表明, 提出的方法针对跨场景缺陷识别 问题优于现有的域适应方法。 [ABSTRACT FROM AUTHOR]