基于卷积神经网络的光学遥感图像检索

提出了一种基于深度卷积神经网络的光学遥感图像检索方法。首先,通过多层卷积神经网络对遥感图像进行卷积和池化处理,得到每幅图像的特征图,抽取高层特征构建图像特征库;在此过程中使用特征图完成网络模型参数和Softmax分类器的训练。然后,借助Softmax分类器在图像检索阶段对查询图像引入类别反馈,提高图像检索准确度,并根据查询图像特征和图像特征库中特征向量之间的距离,按相似程度由大到小进行排序,得到最终的检索结果。在高分辨率遥感图像数据库中进行了实验,结果显示:针对水体、植被、建筑、农田、裸地等5类图像的平均检索准确度约98.4%,增加飞机、舰船后7类遥感图像的平均检索准确度约95.9%;类别信息的引入有效提高了遥感图像的检索速度和准确度,检索时间减少了约17.6%;与颜色、纹理、词袋模型的对比实验表明,利用深度卷积神经网络抽取的高层信息能够更好地描述图像内容。实验表明该方法能够有效提高光学遥感图像的检索速度和准确度。

Optical remote sensing image retrieval based on convolutional neural networks

A method for remote sensing image retrieval based on convolutional neural networks was proposed. First, the convolution and pooling of remote sensing images were conducted by multi-layer convolutional neural networks. The feature maps of each image were obtained, and the high-level features were extracted to build the image feature database. In this process, the training of networks' parameters and the Softmax classifier were completed using feature maps. Then, in the image retrieval stage, classification was introduced by the softmax classifier which will improve the accuracy of image retrieval. Lastly, the remote sensing image retrieval was sorted based on the similarity between the query image and database. Retrieval experiments were performed on the high-resolution optical remote sensing images. The average retrieval precision on five kinds including water, plant, building, farmland and land is 98.4%, and the retrieval precision on seven types (adding plane and ship) is 95.9%. The introduction of class information improves the retrieval precision and speed, saving time by 17.6% approximately. The proposed method behaves better than the methods that based on color feature, texture feature and the bag of words model, and the results show that the high-level feature from deep convolutional neural networks can represent image content effectively. Experimeat indicates that retrieval speed and accuracy of optical remote-sensing images can be effectively increased in this method.