矿物分析在地质勘测及工程应用中都是一项极为关键的工作,在矿物分析中,相比于传统的物理方法和化学方法,拉曼光谱检测能提供更快速的定性定量分析,最重要的是,它对矿物的损伤可以忽略不计。而基于拉曼光谱的数据分析,传统的机器学习方法效果并不理想,尤其在矿物种类极其庞大的情况下。为此,提出一种基于Siamese网络的相似性学习方法,采用Hungarian算法来优化负样本,与传统的机器学习算法相比,得到了鲁棒性最好的结果。Siamese网络计算出矿物之间的相似度,除了能对矿物进行识别,还可以在一定程度上为该矿物的替代材料提供参考。

Mineral identification is a critical task in geological surveys and in many engineering applications. Compared to the physical methods and chemical methods, Raman spectroscopy provides a faster qualitative and quantitative analysis, and most importantly, its damage to the original mineral is negligible in mineral analysis. But the data analysis based on Raman spectroscopy, results of traditional machine learning methods do not work well, especially in the case of minerals with a large categories. This paper proposes a similarity learning method based on Siamese network. After optimizing the negative samples by Hungarian algorithm, and compared with traditional models, we achieve the best robust results. What Siamese network computes is the similarity between minerals, in addition to the identification of minerals, it can also provide a reference for the alternative materials of the mineral to some extent.