Research On Remote Sensing Image Target Detection Based On Deep Neural Network

As a basic application of image recognition,object detection is to detect certain kinds of objects and locate them on a given image.Object detection of remote sensing images is an important component of object detection,which aims to deal with high resolution remote sensing images with large sizes and detect objects such as airplanes,buildings,roads and harbors.Object detection of remote sensing images could be researched for military and civil use,which including military guidance,target tracking,urban planning,disaster monitoring and autopilot.But there still remain many difficulties in these researches because of the large size,tiny targets,clustered backgrounds and obscure distinction of colors on remote sensing images.Deep neural networks could extract multi-level features of remote sensing images automatically and be an auxiliary to object detection,but prevalent object detection framework using deep neural networks could not be used on remote sensing images directly because of the lacking labeled training samples of remote sensing image processing.Object detection methods of remote sensing images using deep neural networks are explored in this paper,based on existing researches and the theoretical basis of deep neural networks.The training of deep neural network models would be accomplished on less labeled samples.The innovation of this paper is discussed from three aspects.(1)Deep neural networks are introduced to the field of classification of remote sensing images,which is the basic task for object detection.Firstly,a new dataset is proposed to extend the training set of remote sensing images,and deep models could be fully trained on the extended dataset.Secondly,four deep neural networks are used for classification,in which deep models are trained by the method of transfer learning.The experimental results demonstrate that deep models could extract features robustly and obtain high accuracy on remote sensing image datasets.Classification is the critical step of object detection and is the primary application of deep neural networks.So the research of classification is the base and premise of the research of object detection.(2)An object detection framework of remote sensing images using deep neural networks is proposed,which uses fully convolutional networks and convolutional neural networks.The proposed framework is researched,trained and tested on airplane detection datasets,which are common in the field of object detection of remote sensing images.Also,new dataset of airplane detection is proposed to extend existing dataset.Firstly,fully convolutional network is used to coarsely segment input image,from which proposals are extracted.Secondly,convolutional neural network is used to classify proposals and obtain bounding-boxes.Thirdly,bounding-box fine-tuning algorithm is proposed to remove overlapped boxes and fine-tuning the location of remaining boxes,which make up the final detection results.Simplified weakly supervised training method is used to train deep models,which reduces the demand of labeled remote sensing image samples.The proposed framework obtains good detection results on both two dataset.(3)A new object detection framework of remote sensing images is proposed based on researches in this paper,which accelerate the detection process significantly.The new framework coarsely segment input image with a fully convolutional network,independent threshold is used to remove regions with low scores.Then,bounding-box generation algorithm is used to obtain bounding-boxes from segment maps directly and a selecting algorithm is used to lessen false positives.At last,bounding-box fine-tuning algorithm is used to adjust the location of boxes and fuse overlapped boxes.The new framework only extracts features once from input images and could improve the detection speed while obtaining precise detection results.Therefore,the new framework could achieve detection tasks with high real time requirements.