Remote Sensing Image Object Detection And Recognition Based On Convolutional Neural Network

With the rapid development of remote sensing technology,especially the emergence of remote sensing satellites,there are a huge number of remote sensing data been produced,which are valuable resource for research.And remote sensing object detection and recognition has been a popular research field since they can be applied to various fields.Remote sensing object detection,finding objects with lable and location in a remote sensing image.And remote sensing object recognition means annotating an object with fine classification.These two tasks have long been concerned in the field of remote sensing.Convolutional Neural Network(CNN)has made remarkable achievements in the field of computer vision since its high-level semantics features.And CNN has been widely used in remote sensing object detection and recognition in recent years.Nevertheless,the existing approaches of CNN for remote sensing object detection rely on huge bounding box data(location information data),which result in a huge cost of manual tagging and a long time for training.Meanwhile,the number of remote sensing objects is too small to support the large scale training.And the existing approaches of CNN for remote sensing object recognition only use high-level semantics features from last convolutional layer,but ignore shallow features of CNN.To address above issues,we conducted a research on remote sensing object detection and recognition based on CNN.Fristly,we will make a brief introduction of research background and current research state for CNN and remote sensing object detection and recognition,especially the analysis and introduction of their research progess.And then,for remote sensing object detection task,we propose a method based on deep feature,which improves mAP(mean Average Precision),reduces mMR(mean Miss Rate)and training without any bounding box data.For remote sensing object recognition task,we design a new CNN model for object recognition and propose an approach to combine different features of convolutional layer,which makes full use of CNN's features to improve the performance of recognition.The main contributions of this dissertation are as follows:(1)At present,the methods of remote sensing object detection baesd on CNN always rely on huge bounding box data to train a network,while the number of remote sensing objects is too small to support this large scale training.In this dissertation,we adopt high resolution remote sensing image to train a CNN,and using this CNN to extract deep feature from low resolution remote sensing to get regions of interests(ROIs),then we adopt multiscale CNNs to confirm these ROIs.Therefore,our method can effectively find object's location and confirm it.The experiments show that,when compared to other similar approaches,our method reaches the best mAP,mMR,and training without any bounding box data.(2)The methods of remote sensing object recognition baesd on deep learning always only use high-level semantics features from last convolutional layer,but ignore shallow features of CNN.Therefore,we designed a new CNN model for object recognition,and propose an end to end CNN model with combining different features of convolutional layer through global average pooling(GAP),which makes full use of CNN's features to improve the performance of recognition.Through the analysis of the comparison with other methods,the proposed method achieves the highest accuracy.