基于自蒸馏的热轧带钢表面缺陷识别

Surface defect recognition of hot-rolled strip steel based on self-knowledge distillation

表面缺陷检测是热轧带钢生产过程中十分重要的环节,是提高热轧带钢产品质量的关键。首先,针对当前带钢表面缺陷图像形貌多样、干扰多以及检测算法精度有待提高的问题,设计了一种自蒸馏框架(Self-knowledge Distillation,SD)来提高模型的缺陷检测精度,通过采用几何变换、色彩增强和图像线性插值的数据增强方法,增加样本数据的表征信息;从标签的概率分布和接近程度两方面提出了新的损失函数(LCK Loss),使模型能更好地学习样本提供的知识,实现相同数据不同表征间的信息传递,进而提高模型的泛化能力。其次,将Poolformer网络应用到热轧带钢表面缺陷检测中,针对Poolformer12网络参数量多的问题,设计了轻量化网络LRAM-Poolformer8,通过减少网络层数和计算量,实现模型加速的目的。最后,在武钢CSP机组的8类带钢表面缺陷数据集上进行了试验,结果表明,所提SD-LRAM-Poolformer8模型的平均识别精度为98.20%,相较于Poolformer12,检测精度提高了1.62个百分点,且计算量仅为原来的56.4%,这证明了新模型在热轧带钢表面缺陷检测中的可行性与有效性。

Surface defect detection is critical aspect in the production process of hot-rolled strip steel,serving as a key factor in improving the quality of these steel products.Firstly,to address the challenges posed by the diverse morphologies of surface defect images,various interferences,and the need for improved detection accuracy,a self-knowledge distillation framework(SD)was designed to enhance the defect detection accuracy of the model.The representational information of sample data was increased by employing data augmentation methods such as geometric transformations,color enhancement,and linear interpolation.Additionally,a novel loss function(LCK Loss)was proposed,considering both the label's probability distribution and closeness,enabling the model to better learn the knowledge provided by the samples and facilitating the transfer of information between different representations of the same data,thereby improving the model's generalization ability.Secondly,the Poolformer network was applied to the surface defect detection of hot-rolled strip steel.To address the issue of a large number of parameters in the Poolformer12 network,a lightweight network called LRAM-Poolformer8 was designed,achieving model acceleration by reducing network depth and computational complexity.Finally,experiments were conducted on 8-class surface defect dataset from WISCO's CSP unit.The results demonstrate that the proposed SD-LRAM-Poolformer8 model achieves average recognition accuracy of 98.20%.Compared to Poolformer12,it shows an improvement of 1.62 percent points in detection accuracy while reducing the computational complexity to only 56.4%of the original model.These findings highlight the feasibility and effectiveness of the new model in surface defect detection of hot-rolled strip steel.