Vehicle Detection In Remote Sensing Imagery Based On Super-resolution Transfer Learning

A large number of remote sensing images have been produced with the rapid development of remote sensing imaging technology.Object detection in remote sensing images such as vehicle detection has been an important research topic in many applications,such as public security,intelligent transportation and urban planning.The lack of annotation information,apparent features,noise interference and other factors in large-scale satellite remote sensing images,have seriously affected the direct application of mainstream vehicle detection algorithms.It has also resulted in difficulty of directly training robust vehicle detectors on satellite remote sensing images with insufficient supervised information.The main contents of this thesis include the following three parts:(1)To overcome the issue of unobvious features in satellite remote sensing images,a super-resolution algorithm based on random forest is introduced,which enhances the feature and reconstruct image.To process remote sensing images with complex background,the proposed super-resolution algorithm combines random-forest based sparse coding with anchor-neighborhood(ANR)scheme into a local linear regression problem.(2)To overcome the issue of noise interference in aerial remote sensing images,a deep-learning model GoogLeNet is introduced to train a robust vehicle detector.By introducing context information to the training data,the trained deep network is more suitable for the vehicle detection under complex background of aerial image.(3)Combing the above two points,this thesis further proposed an integrated framework for vehicle detection in remote sensing images via super-resolution transfer.The framework can be divided into offline part and online part.The offline part is mainly for training the robust detection model on the aerial images.The online part is mainly for the reconstruction of high-resolution satellite images,and then send the reconstruction results into deep convolutional network GoogLeNet for detection.In such a way,the difficult of detection vehicles in low-resolution(target)domains is transformed into the detection in the high-resolution(source)domain.Experiments show that,by comparing to the vehicle detection model trained directly on the satellite remote sensing images,the proposed method can improve the detection performance with a significant gain,which achieves the state-of-the-art results in large-scale dataset.