弱监督深层神经网络遥感图像目标检测模型

Object detection models of remote sensing images using deep neural networks with weakly supervised training method

本文提出了一种遥感图像目标检测框架,克服了遥感图像中由于目标较小且背景复杂造成的目标检测任务中的困难.所提框架包含两种深层神经网络模型,分别是全卷积网络模型和卷积神经网络模型.首先,使用全卷积网络提取遥感图像中可能存在待检测目标的候选区域,避免了对图像的穷举搜索.其次,使用深层卷积神经网络对候选区域分类,通过提取高层特征提高分类正确率.然后,提出了新的遥感图像目标检测数据集,模型的训练全部使用图像级的标签,提出简化弱监督训练方法解决遥感图像目标检测领域目标级标签缺乏的问题.最后,提出一种候选框融合算法,合并重叠候选框的同时调整候选框的位置.提出的模型在本文所提数据集satellite aircrafts dataset和公开数据集aircrafts dataset上进行了测试.实验结果表明,提出的目标检测框架和其他使用深层神经网络的框架相比提高了目标检测的正确率,并具有更高的检测效率.

An object detection framework of remote sensing images is proposed in this paper.It aims to bridge the difficulties in the object detection task of remote sensing images caused by small targets and complex backgrounds.The proposed framework consists of two deep neural network models： fully convolutional network model and convolutional neural network model.First,the fully convolutional network model is used to extract proposals that may contain targets to be detected in the remote sensing images,thus avoiding an exhaustive search in the whole image.Next,the convolutional neural network model is used to classify the proposals.Highlevel features are extracted to improve the rate of classification accuracy.Subsequently,a new dataset of remote sensing images used for object detection is provided.Image-level annotations are used to train all the models in the proposed framework.Simplified weakly supervised training method is used to solve the problem unachievable by object-level annotations in the field of object detection on remote sensing images.Finally,a novel proposal fusing algorithm is proposed,by which the positions of proposals are adjusted while overlapped proposals are fused.The proposed framework is tested on the proposed satellite aircraft dataset and the public aircraft dataset.The experimental results demonstrate that the proposed object detection framework improves the recognition rate as well as the detection efficiency of object detection when compared with other object detection frameworks using deep neural networks.