【摘要】 针对传统变电站人工视频安全帽检测效率低和错误率高的问题，文中提出了一种优化的高精度Faster-RCNN安全帽检测方法。首先，以RPN为主干网络引入特征金字塔，增强检测算法浅层和深层信息表征能力；接着，引入K-Means++聚类算法优化Anchor参数，提升网络对小目标的检测能力；然后，以ROI Align代替RoI池化，消除量化误差和原始图片与特征图的映射偏差，从而提高检测精度；最后，采用开源安全帽数据集对改进的网络进行训练与测试，并将该模型与YOLOv3、RFBnet和传统Faster-RCNN进行了对比。实验结果表明：优化的Faster-RCNN相比RFBnet、YOLOv3和传统Faster-RCNN模型，将mAP值分别提高了6.81%、9.57%和5.09%,达到了92.43%;检测速度为18 frame/s,同时增强了变电站安全帽高精度识别能力。更多还原

【Abstract】 Aiming at low efficiency and a high error rate of safety helmet detectionin substations by the traditional manual video detection method, a helmet detection method based on the improved high-precision Faster-RCNN is proposed. Firstly, the feature pyramid network module is introduced into the RPN backbone network, so that the representation ability of shallow semantic information and deep semantic information are enhanced. Secondly, the K-Means++ clustering algorithm is introduced to improve the anchor parameters, so that the ability to detect small targets is enhanced.And then, RoI pooling is replaced by ROI Align to eliminate the quantization error and the mapping deviation between original image and feature map, thereby improving the detection accuracy. Finally, the open source dataset of safety helmet is used to train and test the improved network, and the model is combined with YOLOv3, RFBnet and traditional Faster-RCNN. The results show that compared with RFBnet, YOLOv3 and traditional Faster-RCNN models, the improved Faster-RCNN increases the mAP value by 6.81%, 9.57% and 5.09%, which reaches 92.43%, and the detection speed is 18 frame/s, which can enhance the high-precision recognition ability of safety helmets in substations.更多还原