Intelligent ore sorting stands as a pivotal technology in contemporary mining and production. To establish a more efficient general framework for mineral image recognition, this paper proposes to combine the inductive biases of convolutional neural networks (i.e., locality and translation invariance) with the non-local qualities of transformer architecture (i.e., globality and long-range dependencies) and introduce the self-attention mechanism, which leads to a new series model, called OreFormer. In fine-grained coal sorting, OreFormer demonstrated preferred performance across gas coal, coking coal, and anthracite, which are higher than the purely ConvNet or vision transformer. At the same time, OreFormer variant models achieved a preferred tradeoff between classification performance and efficiency (i.e., they can simultaneously maintain higher classification accuracy), less computational complexity, and smaller model size. In addition, OreFormer had excellent discriminative and feature representation capability, which can distinguish accurately mineral particles with minor apparent differences between categories.

智能矿石分选是当代采矿和生产的关键技术。为了建立更有效的矿物图像识别通用框架，本文提出将卷积神经网络的归纳偏差（即局部性和平移不变性）与变压器架构的非局部特性（即全局性和远程依赖性）相结合）并引入自注意力机制，从而产生了一个新的系列模型，称为 OreFormer。在细粒煤分选中，OreFormer 在气煤、炼焦煤和无烟煤方面表现出了较好的性能，高于纯 ConvNet 或视觉变压器。同时，OreFormer变体模型在分类性能和效率（即可以同时保持较高的分类精度）、较低的计算复杂度和较小的模型尺寸之间实现了优选的权衡。此外，OreFormer具有出色的判别和特征表示能力，可以准确地区分类别间明显差异较小的矿物颗粒。