EMS-YOLO-Seg: An Efficient Instance Segmentation Method for Lithium Mineral Under a Microscope Based on YOLO11-Seg

Lithium minerals are essential raw materials for new energy storage systems, and accurate instance segmentation of their microscopic images is crucial for efficient resource exploration and utilization. However, existing segmentation methods face challenges when processing lithium mineral images, including complex texture overlaps, missed detection of small particles, and deployment difficulties on edge devices, making it hard to balance segmentation accuracy with inference speed. To address these challenges, this paper proposes an efficient instance segmentation method based on YOLO11-seg, named EMS-YOLO-seg. First, we designed Multi-Scale Partial Convolution (MSPConv) and integrated it into the C3k2 module. The modified C3k2-MSP module optimizes the model’s receptive field and enhances its multi-scale feature extraction capability. We replaced the PSABlock module with the CBAM attention mechanism, introducing the C2PSA-CBAM module, which strengthens the model’s channel focus and feature extraction abilities. The redesigned Segment-LSCDMSP segmentation head reduces computational complexity and improves detection efficiency. Experimental results on our custom-built lithium mineral microscopic image dataset show that compared to the baseline YOLO11n-seg model, the EMS-YOLO-seg model achieved a 0.8% and 0.8% improvement in mAP50box and mAP50:95box, respectively, and a 1% and 0.7% improvement in mAP50mask and mAP50:95mask. Additionally, the model reduced the number of parameters by 52.1%, FLOPs by 18.6%, model size by 49.4%, and increased FPS by 12.7%. This study provides reliable technical support for accurate instance segmentation of lithium mineral microscopic images and demonstrates strong scene adaptability and promising potential for real-time deployment under industrial environments and resource-constrained scenarios.