电子测量与仪器学报

郭家虎,何磊.DEL-YOLO:低照度轻量级煤矿输送带异物检测[J].电子测量与仪器学报,2025,39(12):289-299

DEL-YOLO:低照度轻量级煤矿输送带异物检测

DEL-YOLO:Low-illumination lightweight object detection forconveyor belts in coal mines

DOI：

英文关键词:coal mine conveyor belt foreign object detection deformable convolution lightweight detection head

基金项目:

作者	单位
郭家虎	安徽理工大学电气与信息工程学院淮南232001
何磊	安徽理工大学电气与信息工程学院淮南232001

Author	Institution
Guo Jiahu	School of Electrical and Information Engineering, Anhui University of Science and Technology, Huainan 23200,China
He Lei	School of Electrical and Information Engineering, Anhui University of Science and Technology, Huainan 23200,China

摘要点击次数: 580

全文下载次数: 431

中文摘要:

针对煤矿井下环境中存在的低照度成像质量差、小尺度异物漏检率高以及目标遮挡导致的特征信息缺失等关键问题，提出了一种基于YOLOv11s的低照度轻量化异物检测模型DEL-YOLO。首先，在图像预处理阶段，引入对比度受限自适应直方图均衡化算法来强化低照度图像细节特征，有效提升暗区异物可见性；其次，在网络架构层面创新性地设计了特征提取模块DE-Block，并通过构建DE-C3K2模块，对形状不规则以及存在重叠遮挡等特点的异物进行特征提取；进一步地，在颈部网络嵌入特征融合模块EFC，其通过层间特征相关性强化机制抑制冗余特征融合，同时强化小目标特征表达能力；最后，设计轻量化检测头L-Detect，利用颈部特征共享策略实现参数量压缩。实验结果表明，DEL-YOLO平均检测精度可达80.8%，与YOLOv11相比，平均精确率提升了4.9%，模型的计算量下降了40.74%，参数量下降了41.75%，模型大小仅为6.45 MB。改进模型在显著降低复杂度的同时，仍能有效解决煤矿井下低照度复杂环境中小目标漏检与遮挡目标检测问题。

英文摘要:

Addressing the key issues in coal mine underground environments, such as poor imaging quality under low illumination, high miss-detection rates for small-scale foreign objects, and feature information loss caused by object occlusion, this paper proposes a low-illumination lightweight foreign object detection model, DEL-YOLO, based on YOLOv11s. Firstly, in the image preprocessing stage, the Contrast-Limited Adaptive Histogram Equalization algorithm is introduced to enhance the detailed features of low-illumination images, effectively improving the visibility of foreign objects in dark areas. Secondly, at the network architecture level, an innovative feature extraction module, DE-Block, is designed, and a DE-C3K2 module is constructed to extract features from foreign objects with irregular shapes and overlapping occlusions. Furthermore, a feature fusion module, EFC, is embedded in the neck network, which suppresses redundant feature fusion through an interlayer feature correlation enhancement mechanism while strengthening the feature representation capability for small objects. Finally, a lightweight detection head, L-Detect, is designed, which achieves parameter compression through a neck feature sharing strategy. Experimental results show that DEL-YOLO achieves an average detection accuracy of 80.8%. Compared with YOLOv11, it improves the average precision rate by 4.9%, reduces the computational load by 40.74%, and decreases the number of parameters by 41.75%, with a model size of only 6.45 MB. While significantly reducing complexity, the improved model can still effectively address the issues of small object miss-detection and occluded object detection in the complex low-illumination environments of coal mines.

查看全文查看/发表评论下载PDF阅读器