基于深度学习的煤矸石检测算法

Coal Gangue Detection Algorithm Based on Deep Learning

摘要: 在复杂的煤矸石分拣环境中，传统机器视觉算法易受环境影响、煤矸石之间相互重叠和遮挡、以及分拣背景环境复杂多样。针对上述问题，该研究提出了一种基于改进YOLOv11n的煤矸石目标检测算法。将HGNetV2网络作为主干网络，采用归一化的方式与共享卷积结构，设计轻量化检测头网络，减小模型体积；然后使用特征增强网络BIFPN模块，通过多尺度特征融合来弥补引入轻量级网络带来的检测精度损失，实现在保证检测精度的情况下完成模型轻量化；最后采用Slideloss损失函数，提升分类边界框回归能力以及对匮乏样本的检测精度。试验结果表明，改进的YOLOv11n的参数量、浮点数和模型大小为2.19 × 10⁶、6.5GFLOPS、4.95 M，查准率达到92.6%，召回率达到87.9%，F1得分为90.1。在煤矸识别方面，改进的YOLOv11模型通过优化，在保持最高查准率的同时，达成了最低的参数量和浮点运算次数。这一轻量化特性使其在识别精度与速度上获得了良好的均衡。经过矿场实地实验的充分验证，该模型为煤矸的精确辨识提供了一个强有力的工具。

Abstract: In the complex sorting environment of coal gangue, traditional machine vision algorithms are easily affected by environmental factors, overlap and occlusion between coal gangue, and the diversity of sorting backgrounds. To address these issues, this study proposes a coal gangue object detection algorithm based on improved YOLOv11n. The HGNetV2 network is used as the backbone, along with a normalized approach and shared convolutional structure, to design a lightweight detection head network that reduces the model’s size; then the feature enhancement network BIFPN module is used to compensate for the detection accuracy loss caused by the introduction of the lightweight network through multi-scale feature fusion, achieving model lightweight while ensuring detection accuracy; finally, the Slideloss loss function is adopted to improve the boundary box regression ability for classification and the detection accuracy for scarce samples. Experimental results show that the improved YOLOv11n has a parameter count of 2.19 × 10⁶, a floating-point operation count of 6.5GFLOPS, and a model size of 4.95 M, with a precision of 92.6%, a recall rate of 87.9%, and an F1 score of 90.1. In terms of coal gangue identification, the improved YOLOv11 model achieves the lowest parameter count and floating-point operations while maintaining the highest precision rate through optimization. This lightweight feature allows it to achieve a good balance between recognition accuracy and speed. Fully validated through on-site experiments in mines, this model provides a powerful tool for the accurate identification of coal gangue.