基于动态监督知识蒸馏的输电线路螺栓缺陷图像分类

Bolts are widely used fasteners in transmission lines, and their defect images have the characteristics of small intra-class differences and large inter-class differences. Aiming at the problem that large models with high complexity and excellent performance consume a large amount of computing resources when analyzing bolt defect images, knowledge distillation technology is introduced into the classification of bolt defect images in transmission lines, and a classification method for bolt defect images in transmission lines based on dynamic supervised knowledge distillation is proposed: an adaptive weighting method is adopted in the network output layer to improve the accuracy of small models in learning bolt defect labels; attention transfer is carried out in the network hidden layer to improve the expression ability of bolt features of small models; the adaptive weighting method of the network output layer is combined with the attention transfer mechanism of the network hidden layer to fully improve the bolt defect classification ability of small models. Finally, the effectiveness of the large model using the proposed distillation method to guide the training of the small model is verified through a self-built bolt defect image classification dataset. The experimental results show that the classification accuracy of the small model is increased by 2.17%, the difference in classification accuracy between the small model and the large model is only 0.63%, and the number of parameters of the small model is only 7.8% of that of the large model. The research realizes the efficient classification of bolt defects and achieves a balance between accuracy and resource consumption.

螺栓是输电线路中广泛存在的紧固件,其缺陷图像具有类内差异性小、类间差异性大的特性。针对复杂度高且性能优秀的大模型在分析螺栓缺陷图像消耗大量计算资源的问题,将知识蒸馏技术引入到输电线路螺栓缺陷图像分类中,提出了一种基于动态监督知识蒸馏的输电线路螺栓缺陷图像分类方法:在网络输出层采用自适应加权方法,提高小模型学习螺栓缺陷标签的准确性;在网络隐藏层进行注意力转移,提高小模型螺栓特征的表达能力;将网络输出层的自适应加权方法与网络隐藏层的注意力转移机制相结合,以充分提高小模型的螺栓缺陷分类能力。最后通过自建螺栓缺陷图像分类数据集验证了大模型利用所提蒸馏方法指导小模型训练的有效性,实验结果表明:小模型的分类准确率提高了2.17%,小模型与大模型的分类准确率只差0.63%,且小模型的参数量仅为大模型参数量的7.8%。研究实现了螺栓缺陷的高效分类,达到了精度与资源消耗的平衡。