基于YOLOv5s−SDE的带式输送机煤矸目标检测

Coal gangue target detection of belt conveyor based on YOLOv5s-SDE

  • 摘要: 传统的煤矸图像检测方法需要人工提取图像特征,准确率不高,实用性不强。现有基于改进YOLO的煤矸目标检测方法在速度和精度方面有所提升,但仍不能很好地满足选煤厂带式输送机实时智能煤矸分选需求。针对该问题,在YOLOv5s模型基础上进行改进,构建了YOLOv5s−SDE模型,提出了基于YOLOv5s−SDE的带式输送机煤矸目标检测方法。YOLOv5s−SDE模型通过在主干网络中添加压缩和激励(SE)模块,以增强有用特征,抑制无用特征,改善小目标煤矸检测效果;利用深度可分离卷积替换普通卷积,以减少参数量和计算量;将边界框回归损失函数CIoU替换为EIoU,提高了模型的收敛速度和检测精度。消融实验结果表明:YOLOv5s−SDE模型对煤矸图像的检测准确率达87.9%,平均精度均值(mAP)达92.5%,检测速度达59.9帧/s,可有效检测煤和矸石,满足实时检测需求;与YOLOv5s模型相比,YOLOv5s−SDE模型的准确率下降2.3%,mAP提升1.3%,参数量减少22.2%,计算量下降24.1%,检测速度提升6.4%。同类改进模型对比实验结果表明,YOLOv5s−STA与YOLOv5s−Ghost模型的检测精度明显偏低,YOLOv5s−SDE模型与YOLOv5s模型及YOLOv5s−CBAM模型的检测效果整体相近,但在运动模糊和低照度情况下,YOLOv5s−SDE模型整体检测效果更优。

     

    Abstract: Traditional coal gangue image detection methods require manual extraction of image features. The methods have low accuracy and practicality. The existing coal gangue target detection methods based on improved YOLO have improved in speed and precision, but they still cannot meet the real-time intelligent coal gangue sorting needs of belt conveyors in coal preparation plants. In order to solve the above problems, an improvement is made to the YOLOv5s model, and a YOLOv5s-SDE model was constructed. A method for coal gangue target detection of belt conveyors based on YOLOv5s-SDE is proposed. The YOLOv5s-SDE model enhances useful features, suppresses useless features, and improves the detection effect of small target coal gangue by adding squeeze-and-excitation (SE) module to the backbone network. The model replaces ordinary convolutions with depthwise separable convolutions to reduce parameter and computational complexity. The loss function of the bounding box regression CIoU is replaced by the EIoU. This improves the convergence speed and detection precision of the model. The results of the ablation experiment show that the YOLOv5s-SDE model has a detection accuracy of 87.9% for coal gangue images, a mean average precision (mAP) of 92.5%, and a detection speed of 59.9 frames/s. It can effectively detect coal and gangue, meeting real-time detection requirements. Compared with the YOLOv5s model, the accuracy of the YOLOv5s-SDE model decreases by 2.3%, the mAP increases by 1.3%, the number of parameters decreases by 22.2%, the calculation amount decreases by 24.1%, and the detection speed increases by 6.4%. The comparative experimental results of similar improved models show that the detection precision of YOLOv5s-STA model and YOLOv5s-Ghost model is significantly lower. The detection performance of the YOLOv5s-SDE model, YOLOv5s model and YOLOv5s-CBAM model is generally similar. But in the case of motion blur and low lightning, the overall detection performance of the YOLOv5s-SDE model is better.

     

/

返回文章
返回