基于自训练的多标签岩矿石薄片分类方法

Multi-label rock ore slice classification approach based on self-training

岩矿石薄片识别是一项专业性要求极高的任务,人工识别常出现不可避免的主观错误,且效率极低。深度学习图像识别技术是可以高效进行岩矿石薄片识别的方法,但训练深度学习模型需要大量标注数据,因此如何高效利用有限标注数据具有重要意义。通过采用多标签分类方法,在有标签数据集上先训练一个分类器,然后使用该分类器为大量无标注的岩矿石薄片生成伪标签,最后使用有标签的训练数据和所有无标签数据重新训练模型。结果表明,采用多标签分类方法识别岩矿石薄片结构及矿物是可行的,同时使用半监督学习方法训练模型,在不进行大量人工标注的情况下,可提高该模型的泛化能力。

Rock ore slice identification is a task that requires a high level of expertise.Manual identification often results in unavoidable subjective errors and is highly inefficient.Deep learning image recognition technology can efficiently perform rock ore slice identification,but training deep learning models requires a large amount of annotated data.Therefore,it is important to find efficient ways to utilize limited annotated data.By adopting a multi-label classification approach,a classifier can be trained on a labeled dataset,and then this classifier is used to generate pseudo-labels for a large number of unlabeled rock ore slice images.Finally,the model is retrained using the labeled training data and all the unlabeled data.The results show that the use of multi-label classification approach for identification of rock ore slice structures and minerals is feasible.Additionally,this paper employs a semi-supervised learning approach to train the model and improve the model s generalization ability without requiring a large amount of manual annotation.