基于EMD距离度量的小样本矿物图像分类

Few-shot Mineral Image Classification Based on EMD Metric Learning

在复杂的地质勘探条件下准确完成矿物识别是一项重要的任务。基于数据驱动的深度学习模型能精确识别各类岩石矿物,但需要构建庞大且完备的数据集,在实际情况下难以应用。针对此问题,结合小样本学习、度量学习以及元学习训练策略,使用EMD距离度量计算图像之间的结构距离,构建一种适应于小样本矿物的图像分类模型。核心思想在于利用图块级别度量并引入交叉参考权重机制,有效减少同类差异大和背景杂乱带来的影响,优于图与图判定分类的模型。在mini-ImageNet数据集上,5-way 1-shot和5-way 5-shot设置的分类准确率分别提高至55.91%、67.58%;将算法应用于小样本黏土矿物数据集上,5-way 5-shot设置的分类准确率为92.65%。实验结果表明,利用度量学习方法的分类精度高于其他小样本学习方法。

Mineral identification is an important task of geological survey,which could be a big challenge in complex geological conditions.Data-driven deep learning model could accurately identify all kinds of rocks and minerals,but large and complete datasets should be constructed,and it was difficult in practical situations.Aiming at the problems,herein,an image classification model adapted to small sample minerals was proposed.It adopted the earth mover′s distance(EMD)as a metric to calculate the structural distance between images,which combined the small sample learning,the metric learning and metalearning training strategy.The core idea of the method was to use the measure of graph block level and introduce the cross-reference weight mechanism,which could effectively reduce the influence caused by the large difference of the same class and the clutter of the background,which was better than the model of graph to graph direct decision classification.The results of 5-way 1-shot and 5-way 5-shot classification experiment on mini-ImageNet showed that the classification accuracy of 5-way 1-shot and 5-way 5-shot settings was improved to 55.91%and 67.58%.When the algorithm was applied to a Few-shot clay mineral data-set,the classification accuracy in 5-way 5-shot setting reached 92.65%.The experimental results showed that the classification accuracy of metric learning method was higher than other few-shot methods.