摘要
针对随着城市化的快速发展,城市与城市间的辨识度越来越弱,城市地标的概念越来越热门这一现象,提出了一种基于深度学习的建筑物识别方法;使用改进的Faster R-CNN算法作为训练模型,首先,将需要识别的图片输入深度神经网络,提取出特征框图;然后,模型通过区域建议网络预测目标建筑物所在位置的区域建议,并将这些区域建议映射到特征框图上,RoI Pooling层将这些区域建议转换成固定大小的特征框图;最后使用非极大值抑制从预测边界框中移除相似的结果,得到预测边界框以及边框中目标建筑物的类别和概率;实验结果表明:在训练数据集充足的条件下,使用此方法对地标建筑物的识别率能达到90. 8%,通过与其他模型比较分析,该模型具有较好的识别效果。
Aiming at the rapid development of urbanization,the identification between cities and cities is getting weaker and weaker,and the concept of urban landmarks is becoming more and more popular,This paper proposes a building recognition method based on deep learning. Using the improved Faster R-CNN algorithm as the training model,first,we input the image to be identified into the CNN network and extract the feature block diagram. Then,the model predicts the regional recommendations of the location of the target building through the RPN network,and maps these regional recommendations to the feature block diagram,the RoI Pooling layer converts these regional recommendations into fixed-size feature blocks. Finally,we use non-maximum suppression to remove similar results from the prediction bounding box to get the predicted bounding box,the category and probability of the target building in the border. The experimental results show that the recognition rate of landmark buildings can reach 90. 8% under the condition that the training data set is sufficient. Compared with other models,the model has a good recognition effect.
作者
邓瑞
林金朝
杨宏志
DENG Rui;LIN Jin-zhao;YANG Hong-zhi(School of Communication and Information Engineering,Chongqing University of Postsand Telecommunications, Chongqing 400065,China)
出处
《重庆工商大学学报（自然科学版）》
2019年第4期17-22,共6页
Journal of Chongqing Technology and Business University:Natural Science Edition
基金
国家自然科学基金(61301124,61471075,61671091)
重庆科委自然科学基金(CSTC2016JCYJA0347)
重庆高校创新团队建设计划(智慧医疗系统与核心技术)
关键词
深度学习
建筑物识别
城市化
deep learning
building identification
urbanization
