Studies On Dense Small Object Detection In Remote Sensing Images

Object detection is one of the core concerns in the field of computer vision.According to different applicable scenarios and targets,object detection has developed a large number of research sub-fields.Among them,with the development of aerospace technology and the widespread application of unmanned aerial vehicle in various industries,the utilize of information in airspace image has become the focus of research,especially image research work based on visible spectral remote sensing images,moreover,have found its important application value in production,life living and military activities,it has also become a hot spot for researchers.Remote sensing object detection,as one of the methods to extract the core information from remote sensing images,has attracted the attention of many researchers.However,due to the specific performance of remote sensing images,it cannot be directly applied to general object detection methods.It is necessary to develop a highly customized structure and method to solve the influence and detection error caused by its particularity.To this end,this thesis first reviews the main development process and design ideas of general object detection,analyze the particularity of remote sensing targets and remote sensing data sets,Coupled with the use of deep learning theory and Convolutional Neural Network as the fundamental,and research on remote sensing image object detection methods,which can be reflected in the following works:（1）Developed a modified cascade convolutional remote sensing object detection network,by analyzing the characteristics of the targets in the remote sensing images,combined with the two-stage method struct of general object detection as the basic structure,adding a feature pyramid network that can effectively extract multi-scale features,using deep residual network Res Net-50 as backbone network,deformable convolution is applied to improve the feature extraction process,and a three-level cascade structure is used in the final object detection network to avoid the occurrence of over-fitting in training process,and at the same time improve the detection Intersection over Union threshold.This enables the modified network to more comprehensively use multi-scale information in remote sensing images for detection,as well enhances the ability to detect small objects.（2）Based on the additional information in remote sensing images recorded as GSD（Ground Sample Distance）,a super-resolution remote sensing object detection method with GSD information is researched and proposed.The main idea is to use the characteristics of remote sensing image coverage and rich regional texture to associate the image complexity estimation method with GSD,and introduce it as additional information to the object detection network.At the same time,the super-resolution network combined with GSD is used to perform image processing Selectively.Finally achieves that small objects that are difficult to be detected in an image with a large GSD are resized to a more appropriate scale,and the detection efficiency and accuracy of the small object dense area in remote sensing image are improved.