摘要
为提高煤矸石分拣的精度和可靠性,提出了一种基于改进YOLO v4的煤矸石识别网络,引入了Focal损失函数,使用K-means++聚类算法优化初始锚定框,将PANet中的五次卷积操作替换为CSP结构,同时引入空洞卷积的金字塔结构,降低模型参数,实现模型的轻量化,增加了一条跨连接边构成BiFPN结构,提高对中等目标的检测能力,得到My-YOLO v4目标检测模型。本研究对所提出的My-YOLO v4识别检测方法与SSD、YOLO v3、YOLO v4三种检测方法进行实验对比分析。实验结果表明,该检测算法在测试集上检测煤与煤矸石混合的mAP值为98.14%,FPS为28.3 f/s,相较于SSD、YOLO v3检测算法识别精度分别提高了5.41%、2.87%,相较于YOLO v4目标检测模型识别速度提高了7.7 f/s,通过对比分析实验数据验证了My-YOLO v4目标检测模型整体性能的有效提高。
In order to enhance the accuracy and reliability of coal gangue recognition, a coal gangue recognition network based on improved YOLO v4 is proposed, the initial anchoring frame is optimized using K-means++ clustering algorithm, the five convolution operations in PANet are replaced with CSP structure, and the pyramid structure of hole convolution is introduced at the same time to reduce the model parameters and realize the model light weight, adding a cross-connected edge to form a BiFPN structure to improve the detection capability of medium targets, and obtaining the My-YOLO v4 target detection model. The proposed My-YOLO v4 recognition detection method is compared and analyzed with three detection methods, SSD, YOLO v3 and YOLO v4, by collecting mixed samples of coal and gangue in the field and using relevant experimental equipment. The experimental results show that detection algorithm detects coal mixed with gangue on the test set with mAP value of 98.14% and FPS of 28.3 frames/second, which improves the recognition accuracy by 5.41% and 2.87% compared with SSD and YOLO v3 detection algorithms, respectively, and improves the recognition speed by 7.7 frames/second compared with YOLO v4 target detection model, by comparing The analysis of experimental data verifies the effective improvement of the overall performance of My-YOLO v4 target detection model.
作者
崔斌
陈林
亓玉浩
张坤
赵得福
黄梁松
李明霞
孔祥俊
杜明超
蒋祥卿
刘源
CUI Bin;CHEN Lin;QI Yuhao;ZHANG Kun;ZHAO Defu;HUANG Liangsong;LI Mingxia;KONG Xiangjun;DU Mingchao;JIANG Xiangqing;LIU Yuan(Shandong Key Laboratory of Robotics and Intelligent Technology,Shandong University of Science and Technology,Qingdao 266590,China;Opritel Environmental Technology Co.,Ltd.,Qingdao 266200,China;Qingdao Huaxia Rubber Industry Co.,Ltd.,Qingdao 266200,China;Beidou Tiandi Co.,Ltd.,Jining 710000,China;Qinghai Energy Development(Group)Co.,Ltd.,Xining 810008,China)
出处
《煤炭工程》
北大核心
2023年第12期161-166,共6页
Coal Engineering
基金
山东省重大科技创新工程项目(2019SDZY04)。
