摘要
设计了一个浅层卷积神经网络来代替预训练模型中的全连接层,将预训练网络提取的CNN特征作为图像输入设计好的浅层CNN网络,对比微调预训练模型的方法,能够更好地适应航拍图像定位任务。为进一步提高航拍图像的定位准确率,利用无人机航拍图像时间连续的特点,通过在CNN的分类阶段加入Bi-LSTM网络,使网络在分类时能够以多张图像特征作为判断依据。实验表明,时序图像定位方法定位准确率稳定在0.89左右,对比单张图像定位方法准确率提升5%左右。
A shallow Convolution Neural Network （CNN） is designed to replace the fully-connected layer in the pre-training model. The CNN feature extracted by the pre-training network is taken as input image to the designed shallow CNN network. Compared with the fine-tuning method for the pre-training model, this method can better adapt to the aerial image localization tasks. In order to further improve the locating accuracy of aerial images, the Bi-LSTM network is added to the network at the CNN classification stage by use of the temporally-continuous characteristics of the UAV aerial image. Thus the features of multiple images can be taken as the criterion for the network classification. Experiments show that the accuracy of sequential image locating method reaches a stable level at around 0.89, and is improved by about 5% compared with the single-image locating method.
出处
《电光与控制》
北大核心
2017年第12期51-55,66,共6页
Electronics Optics & Control
基金
吉林省自然科学基金(20130101069JC)
关键词
无人机
航拍图像
图像分类
图像定位
预训练网络
CNN
Unmanned Aerial Vehicle （UAV）
aerial image
image classification
image localization
pre-training network
CNN
