Abstract
Mineral identification technology is a critical technology in the construction of smart mines. To enable effective deployment and implementation of rapid mineral sorting for valuable ores on edge computing devices, we propose a lightweight identification method for lithium minerals under visible light microscopy based on YOLOv8, named Minima-YOLO. First, by scaling down and limiting the number of channels in the YOLOv8 backbone, we introduced a smaller network, YOLOv8-tiny. Next, we redesigned a new lightweight feature extraction module, Faster-EMA, using PConv and the EMA attention mechanism, replacing the original C2f module. Third, we incorporated GhostConv, a cost-effective downsampling method, as a replacement for standard convolutions. Finally, to mitigate the impact of deeper backbone network layers, we introduced the Slim-Neck structure in the Neck, further reducing the model size. Ultimately, Minima-YOLO achieves a 99.4% mAP50 on our self-constructed lithium mineral image dataset, with FLOPs reduced to 2.3 G, parameters to 0.72 M, and a model size of just 1.63 MB, while maintaining an FPS of 103. A series of comparative experiments confirmed its superior performance over other advanced object detection algorithms. This algorithm, with its highly efficient lightweight design and rapid inference speed, provides an intelligent, efficient, and eco-friendly computer vision method for the rapid sorting of lithium mineral components.