基于小样本的传送带异物视频检测算法  

Video detection algorithm based on small sample for foreign objects in conveyor belt

作　　者：贾尚锋 王志强 郝鑫宇 JIA Shangfeng;WANG Zhiqiang;HAO Xinyu(Guoneng Zhunneng Group Co.,Ltd.,Ordos Inner Mongolia 010300,China)

机构地区：[1]国能准能集团有限责任公司,内蒙古鄂尔多斯010300

出　　处：《中国安全科学学报》2023年第S01期134-137,共4页China Safety Science Journal

摘　　要：为解决基于深度学习的传送带异物视频检测算法中实际采集到的传送带异物数据集较少的问题,提出一种基于YOLOX的传送带异物检测算法。首先,使用MobileNet V3替换YOLOX中主干特征提取网络CSPDarknet,以减少主干特征的计算时间;然后,将深度可分离卷积神经网络替代YOLOX中加强特征提取网络中的卷积网络,以减少加强特征提取网络的计算时间;最后,在加强特征提取网络中深度可分离卷积神经网络前添加卷积注意力模块(CBAM)以提高检测准确率。结果表明:经试验验证本文算法在使用少量样本的情况下能够完成训练并实现96.5%的识别准确率;同时本算法的平均时间仅为0.019 s;识别帧率可以达到40 FPS。Since the data sets of foreign objects in the conveyor belt collected by the video detection algorithm based on deep learning for foreign objects in the conveyor belt were fewer,this paper proposed a detection algorithm based on YOLOX for foreign objects in the conveyor belt.First,MobileNet V3 was used to replace the backbone feature extraction network CSPDarknet in YOLOX to reduce the calculation time of the backbone features,and then the convolutional network in the enhanced feature extraction network in YOLOX was replaced with a deep separable convolutional neural network to reduce the calculation time of the enhanced feature extraction network.Finally,a convolutional block attention module(CBAM)was added before the deep separable convolutional neural network in the enhanced feature extraction network to improve detection accuracy.Experiments have verified that the algorithm used in this paper can complete the training and achieve a recognition accuracy of 96.5%when using a small number of samples.At the same time,the average time of this algorithm is only 0.019 s,and the recognition frame rate can reach 40 FPS.

关 键 词：小样本 异物检测 迁移学习 YOLOX 卷积注意力模块(CBAM) 

分 类 号：X936[环境科学与工程—安全科学]