【摘要】 针对台风各阶段尺寸与纹理结构差异较大,存在识别难度大、准确率低的问题,提出改进Faster RCNN的台风云系识别方法。首先为模型选择合适的特征提取网络,然后在原始Faster RCNN基础上优化区域网络参数,改进感兴趣区域池化网络,提高模型检测各阶段台风的鲁棒性。为防止训练过程中正负样本不均衡,引入在线困难样本挖掘机制。实验结果表明,相较于原始Faster RCNN,改进后的Faster RCNN检测平均准确率提高了4.7%,基于ResNet50的Faster RCNN对台风云系识别更好,可以更好地满足台风云系识别需求。更多还原

【Abstract】 Aiming at the large difference in size and texture structure of each stage of typhoon can lead to the problems of the difficulty of recognition and the low accuracy,an improved Faster RCNN identification method of typhoon cloud system is proposed. First,select the appropriate feature extraction network for the model. Then,based on the original Faster RCNN,optimize the Region Proposal Network parameters and improve the region of interest pooling network to improve the robustness of the typhoon at each stage of the model detection,and introduce an online hard example mining mechanism to prevent the imbalance of positive and negative samples in the training process. The experimental results show that the Faster RCNN based on ResNet50 is better. At the same time,compared with the original Faster RCNN,the average detection accuracy of the improved Faster RCNN is increased by 4.7%,which can better meet the requirements of typhoon system identification.更多还原