高如新, 杜亚博, 常嘉浩.基于改进YOLOX-S的轻量化煤矸石检测方法研究[J].河南理工大学学报（自然科学版）,2024,43(4):133-140.

GAO R X, DU Y B, CHANG J H.Study on lightweight coal gangue detection method based on improved YOLOX-S[J].Journal of Henan Polytechnic University(Natural Science) ,2024,43(4):133-140.

基于改进YOLOX-S的轻量化煤矸石检测方法研究

高如新1,2,3, 杜亚博1,2,3, 常嘉浩1,2,3

1.河南理工大学 电气工程与自动化学院，河南 焦作  454000;2.河南省煤矿装备智能检测与控制重点实验室，河南 焦作  454000;3.河南省智能装备直驱技术与控制国际联合实验室，河南 焦作  454000

摘要: 目的 为了探索基于现有机器视觉煤矸石检测方法的模型参数量、计算量对检测速度和嵌入式设备的影响， 方法 提出一种基于改进的无锚框YOLOX-S轻量化煤矸石检测模型。为使模型能提取更真实的煤矸石特征信息，收集分选现场煤矸石样本，保证实际环境下的煤矸石检测效果，适应实际生产环境。结合CSPNet，将输入的特征图分割成两个分支，实现更丰富的梯度组合，同时减少模型计算量；之后在其中一条分支使用Ghost轻量化卷积，通过少量常规卷积生成一组特征图，达到初次减少计算量和参数量的效果，然后在此特征图基础上经过简单线性变化操作，生成一组新的特征图，将两组特征图进行融合，降低对计算资源需求的同时，也达到了常规卷积相同的特征提取效果；引入LeakyReLU激活函数减弱模型梯度消失的影响，提取更深更多的特征信息；最后融合两个分支特征，保证较高的检测精度，提升模型检测速度。采用CIOU Loss（complete IOU loss）优化目标边界框回归损失函数，使模型回归损失收敛更快，提高模型目标定位能力。 结果 与原模型相比，本文改进模型在保证较高的平均精度均值90.51%情况下，模型参数减少47%，计算量减少49%，检测速度达到50 帧/s。 结论 轻量化煤矸石检测模型使智能化煤矸石检测在实际生产环境中具有一定的应用前景。

关键词:煤矸石检测;YOLOX-S;轻量化;目标定位;检测速度

doi:10.16186/j.cnki.1673-9787.2022070020

基金项目:国家重点研发计划项目（2018YFC0604502）

收稿日期:2022/07/10

修回日期:2023/03/17

出版日期:2024/07/15

Study on lightweight coal gangue detection method based on improved YOLOX-S

GAO Ruxin1,2,3, DU Yabo1,2,3, CHANG Jiahao1,2,3

1.School of Electrical Engineering and Automation，Henan Polytechnic University，Jiaozuo  454000，Henan，China;2.Henan Key Laboratory of Intelligent Detection and Control of Coal Mine Equipment，Jiaozuo  454000，Henan，China;3.Henan International Joint Laboratory of Direct Drive and Control of Intelligent Equipment， Jiaozuo  454000，Henan，China

Abstract: Objectives  To explore the influence of the existing machine vision-based coal gangue detection method in the model parametric quantity，the amount of computation on the detection speed and deployment of embedded devices，  Methods a lightweight gangue detection model based on the improvement of the anchor-free YOLOX-S was proposed.By collecting and sorting samples of coal gangue at the site，the model could extract more real coal gangue feature informations to ensure the effect of coal gangue detection in the actual environment and adapt to the actual production environment.Combined with CSPNet，the input feature map was split into two branches to realize richer gradient combinations and to reduce the computational amount of the model；then in one of the branches，Ghost lightweight convolution was used to generate a set of feature maps through a small amount of conventional convolution to reduce the computational amount and the number of parameters for the first time，and then on the basis of this feature map after a simple linear change operation，a new set of feature maps were generated，and finally the two sets of feature maps were combined to form a new set.Then，based on this feature map，a new set of feature maps was generated by a simple linear change operation，and finally，the two sets of feature maps were fused，which reduced the demand for computational resources，and at the same time，achieved the same feature extraction effect as that of the conventional convolution.The LeakyReLU activation function was introduced to alleviate the effect of the vanishing gradient of the model，and extracted more feature informations in a deeper and more.Finally，the features of the two branches were fused，which ensured a higher detection accuracy and improved the detection speed of the model.The CIOU Loss （complete IOU loss） was used to optimize the target bounding box regression loss function，which made the model regression loss converge faster and improved the model target localization ability.  Results Compared with the original model，the improved model ensured a higher average accuracy of 90.51% of the average value of the case，the model parameters were reduced by 47%，the computation amount was reduced by 49%，and the detection speed reached 50 FPS.  Conclusions The lightweight coal gangue detection model made the intelligent coal gangue detection in the actual production environment have a certain application prospect.

Key words:coal gangue detection;YOLOX-S;lightweight;object localization;detection speed