Remote Sensing Image Object Detection Based On Anchor-free And Lightweight

As a basic and challenging visual task,remote sensing image object detection is not only widely used in the military field,such as accurately obtaining battlefield information and conducting precise strikes on the enemy.It also has a wide range of applications in civilian fields,such as environmental management,regional planning,and mineral resource exploration.The existing remote sensing image object detection methods are mostly based on anchor boxes.Such methods will introduce a large number of hyperparameters,increase memory usage and redundant calculations due to too many anchor boxes,and even cause serious problems such as imbalance of positive and negative samples.In addition,the remote sensing image has the characteristics of complex background,large scale variation and dense arrangement in any directions of objects,which further increases the difficulty of object detection in remote sensing image.Besides,the size of current remote sensing images is getting larger and larger,and the amount of data is gradually increasing.In order to obtain better detection performance,the depth and complexity of the model also increase,although this greatly improves the detection accuracy.However,in application scenarios that have high requirements for model complexity,such as the military field,the model needs to be deployed to drones and other equipment for monitoring tasks.The model will not be easy to deploy when it is complex.Building a stable,lightweight,and high-performance anchorfree object detector for remote sensing image is the focus of this research.The main research work of this thesis is as follows:(1)Aiming at the problems caused by the existing remote sensing image object detection following the anchor mechanism,we propose an anchor-free remote sensing image object detection method.In this method,we design a dense path aggregation feature pyramid network,which makes full use of high-level semantic information and low-level texture information to improve the network’s ability to detect objects of different scales.In addition,we also proposed a center region sampling strategy to avoid the influence of fuzzy samples or noise samples on the model performance during the sampling process.Experiments show that the method in this chapter can achieve better performance on remote sensing image object detection tasks.(2)In view of the characteristic that the object in remote sensing image has any direction,we propose a method for directional object detection in remote sensing images based on key points.In this method,considering that the complex background of the remote sensing image leads to a lot of noise information in the features,which results in the features being relatively rough,we propose a semantic transfer block to refine the features.In addition,considering the variety of object sizes in remote sensing images,an adaptive Gaussian heatmap is proposed to locate and classify objects.Experiments show that the method in this chapter can achieve better performance on the task of detecting directional objects in remote sensing images.(3)Aiming at the problems that the remote sensing image object detection model is overparameterized,we propose a remote sensing image object detection method based on knowledge distillation.A lightweight detection network is obtained by this method.Experiments show that this method can greatly reduce the complexity of the model with a small loss of accuracy.