Aiming at the problem of low accuracy of traditional target detection methods for target detection in endoscopes in substation environments, a CNN-based real-time detection method for masked targets is proposed. The m...Aiming at the problem of low accuracy of traditional target detection methods for target detection in endoscopes in substation environments, a CNN-based real-time detection method for masked targets is proposed. The method adopts the overall design of backbone network, detection network and algorithmic parameter optimisation method, completes the model training on the self-constructed occlusion target dataset, and adopts the multi-scale perception method for target detection. The HNM algorithm is used to screen positive and negative samples during the training process, and the NMS algorithm is used to post-process the prediction results during the detection process to improve the detection efficiency. After experimental validation, the obtained model has the multi-class average predicted value (mAP) of the dataset. It has general advantages over traditional target detection methods. The detection time of a single target on FDDB dataset is 39 ms, which can meet the need of real-time target detection. In addition, the project team has successfully deployed the method into substations and put it into use in many places in Beijing, which is important for achieving the anomaly of occlusion target detection.展开更多
目的在基于深度学习的目标检测模型中,浅层特征图包含更多细节但缺乏语义信息,深层特征图则相反,为了利用不同深度特征图的优势,并在此基础上解决检测目标的多尺度问题,本文提出基于卷积核金字塔和空洞卷积的单阶段目标检测模型。方法...目的在基于深度学习的目标检测模型中,浅层特征图包含更多细节但缺乏语义信息,深层特征图则相反,为了利用不同深度特征图的优势,并在此基础上解决检测目标的多尺度问题,本文提出基于卷积核金字塔和空洞卷积的单阶段目标检测模型。方法所提模型采用多种方式融合特征信息,先使用逐像素相加方式融合多层不同大小的特征图信息,然后在通道维度拼接不同阶段的特征图,形成具有丰富语义信息和细节信息的信息融合特征层作为模型的预测层。模型在锚框机制中引入卷积核金字塔结构,以解决检测目标的多尺度问题,采用空洞卷积减少大尺寸卷积核增加的参数量,合理地降低锚框数量。结果实验结果表明,在PASCAL VOC2007测试数据集上,所提检测框架在300×300像素的输入上检测精度达到79.3%mAP(mean average precision),比SSD(single shot multibox detector)高1.8%,比DSSD(deconvolutional single shot detector)高0.9%。在UCAS-AOD遥感数据测试集上,所提模型的检测精度分别比SSD和DSSD高2.8%和1.9%。在检测速度上,所提模型在Titan X GPU上达到21帧/s,速度超过DSSD。结论本文模型提出在两个阶段融合特征信息并改进锚框机制,不仅具有较快的检测速度和较高的精度,而且较好地解决了小目标以及重叠目标难以被检出的问题。展开更多
文摘Aiming at the problem of low accuracy of traditional target detection methods for target detection in endoscopes in substation environments, a CNN-based real-time detection method for masked targets is proposed. The method adopts the overall design of backbone network, detection network and algorithmic parameter optimisation method, completes the model training on the self-constructed occlusion target dataset, and adopts the multi-scale perception method for target detection. The HNM algorithm is used to screen positive and negative samples during the training process, and the NMS algorithm is used to post-process the prediction results during the detection process to improve the detection efficiency. After experimental validation, the obtained model has the multi-class average predicted value (mAP) of the dataset. It has general advantages over traditional target detection methods. The detection time of a single target on FDDB dataset is 39 ms, which can meet the need of real-time target detection. In addition, the project team has successfully deployed the method into substations and put it into use in many places in Beijing, which is important for achieving the anomaly of occlusion target detection.
文摘目的在基于深度学习的目标检测模型中,浅层特征图包含更多细节但缺乏语义信息,深层特征图则相反,为了利用不同深度特征图的优势,并在此基础上解决检测目标的多尺度问题,本文提出基于卷积核金字塔和空洞卷积的单阶段目标检测模型。方法所提模型采用多种方式融合特征信息,先使用逐像素相加方式融合多层不同大小的特征图信息,然后在通道维度拼接不同阶段的特征图,形成具有丰富语义信息和细节信息的信息融合特征层作为模型的预测层。模型在锚框机制中引入卷积核金字塔结构,以解决检测目标的多尺度问题,采用空洞卷积减少大尺寸卷积核增加的参数量,合理地降低锚框数量。结果实验结果表明,在PASCAL VOC2007测试数据集上,所提检测框架在300×300像素的输入上检测精度达到79.3%mAP(mean average precision),比SSD(single shot multibox detector)高1.8%,比DSSD(deconvolutional single shot detector)高0.9%。在UCAS-AOD遥感数据测试集上,所提模型的检测精度分别比SSD和DSSD高2.8%和1.9%。在检测速度上,所提模型在Titan X GPU上达到21帧/s,速度超过DSSD。结论本文模型提出在两个阶段融合特征信息并改进锚框机制,不仅具有较快的检测速度和较高的精度,而且较好地解决了小目标以及重叠目标难以被检出的问题。
