Research On Remote Sensing Image Object Detection Based On Deep Learning

Remote sensing is a method of detecting and analyzing objects and the scene in which it is located using electromagnetic waves without touching the object.Earth observation technology has made great progress in the past decades.The target detection technology of remote sensing images is one of the keys to remote sensing data processing,and has been a hot issue sought after by research scholars.This thesis focuses on a deep learning-based object detection method for remote sensing images,with three major difficulties in the task of object detection in remote sensing scenes,and the main research contents and results are as follows.1.A rotating object detection method based on feature-complete transformation is proposed for the problem of random object angle in remote sensing scenes.The method uses a cascaded single-stage detection method to constitute a two-stage detection method and proposes and presents a feature-complete transformation to expand the detailed features.The first stage of the method uses a horizontal anchor to calculate the initial detection results based on the original multi-size feature map obtained from the feature pyramid.The second step takes the above detection results as a rotating anchor and refines the detection results using the expanded features.The experiment proves to have good detection performance also considering the faster detection speed.2.A few shot detection method based on global semantic analysis is proposed for the problem that small samples of rare categories in remote sensing scenes are difficult to learn quickly.The method uses transfer learning training to improve the reuse of existing knowledge and proposes a global semantic analysis module to further mine the background information before the image,which inherits the meta-learning idea with the ability to quickly adapt to unseen classes.In the training aspect,the global semantic analysis module uses additional semantic segmentation to supervise signal training,so that it focuses on the semantic understanding during feature extraction,which in turn improves the detection performance of small sample scenes.Experiments prove that the method has fast learning capability and good detection performance in few shot scenes.3.A small object detection method based on generative adversarial network is proposed for the problem of difficult detection of small objects with missing information in remote sensing scenes.The method implements end-to-end detection of low-resolution small objects by connecting the generative and detection structures in series.The gradient of the detector during training will also guide the training of the generator,thus shifting the generated objects from obeying the natural distribution as much as possible to being as suitable for detection as possible.For more efficient gradient information transfer,proposed LightGAN,a lightweight generative network.the LightGAN and the detector are considered as a whole in the training phase and trained against the discriminator.It is experimentally demonstrated that the method has better performance and is more suitable for small object detection than the method conducted with super-resolution and detection distributions.