近年来,随着深度学习在地学领域中的应用,矿物图像识别变得越来越重要。虽然已经有研究将深度学习应用于矿物图像识别,并取得了一定的成果,但在大规模矿物数据集上的识别准确率仍然有待进一步提高。不同矿物之间可能存在细微的形态、纹理和颜色差异,符合细粒度识别算法特征,但以往的研究中很少有人采用细粒度方法进行矿物识别。所以本文提出了一种基于Next-ViT模型的细粒度矿物识别方法,通过引入渐进式多粒度训练拼图技术,实现对矿物图像的精确分类。首先采用Next-ViT模型作为特征提取器,该模型结合了Transformer结构和卷积神经网络的优势,能够提取到丰富的图像特征;接下来利用随机拼图生成器创建不同粒度级别的矿物拼图,这些拼图包含从细节到整体的多种信息。训练过程中采用渐进式多粒度训练策略,在训练的早期阶段,模型主要关注细粒度的特征,通过学习拼图中的细节信息来区分不同的矿物,随着训练的深入,模型逐渐将注意力转移到更大粒度级别的特征上,学习更加抽象和全局的信息。通过这种方式,模型能够充分利用不同粒度级别的信息,提高矿物识别的准确性。实验结果表明,该模型在常见的36种矿物数据集上取得了86.5%的准确率,有效地提高了矿物识别的准确率。这表明,细粒度识别方法对于矿物识别是有效的。

In recent years mineral image recognition has become increasingly important for mineral identification with the use of deep learning. While such application has gained some success, further improvement is needed to enhance the identification accuracy on large-scale mineral datasets. The fine differences in morphology, texture, and color between different minerals may align with the characteristics of fine-grained recognition algorithms, yet results of fine-grained recognition for mineral identification have rarely been reported. This paper proposes a fine-grained mineral identification model based on Next-ViT, which allows precise classification of mineral images by progressive model training across different granularity levels. In this approach, Next-ViT, which combines the advantages of transformer and convolutional neural network, is utilized to extract rich image features; a random jigsaw generator is then employed to create mineral puzzles at different granularity levels encompassing various information from detailed to general. The model training involves progressive learning across multiple granularity levels. In the early stages, the model primarily focuses on fine-grained features, learning detailed information from the puzzles to differentiate between different minerals; as training progresses, model learning gradually shifts to higher granularity levels, capturing more abstract and global information. Through this approach, the model can effectively utilize information across multiple granularity levels, thereby improving the accuracy of mineral identification. Experimental results demonstrated the effectiveness of this approach, with the proposed model achieving an accuracy of 86.5% in mineral identification on a dataset on 36 common minerals.