نمونه متن انگلیسی مقاله
Object detection in aerial images is a challenging task which plays an important role in many fields, such as intelligent traffic management, fishery management and so on. Different from object detection in natural images, the orientation of objects in aerial images is arbitrary. The axis-aligned bounding box detection, which is always used in traditional object detection methods, will cover a lot of redundant information and deteriorate the detection results when it is used to locate the object in aerial images. Therefore, traditional object detection methods are no longer applicable for aerial images. In order to promote the object detection performance in aerial images, we propose a novel orientation robust object detection model based on rotated non-maximum suppression (R-NMS). In addition, we adjust the anchor setting according to the diversity shapes of the aerial objects to enhance the performance of the model. Our model is tested on the public DOTA dataset, and the mAP is 16.31% higher than the baseline, indicating that our method is very effective and competitive in the object detection of aerial image.
In recent years, inspired by deep learning, significant progress has been made in object detection. By learning the in-depth presentation of the region of interest (RoI), the deep learning-based detectors make it possible to precisely locate and classify objects on the image. There are many works, such as R-CNN1, Fast R-CNN2, Faster R-CNN3, YOLO4, YOLO90005 and so on, have achieved excellent performance on object detection task in natural scenes. However, these methods are horizontal region-based detection methods and are not suitable for aerial image detection tasks because of the arbitrary rotation characteristics of aerial objects. Using axis-aligned bounding boxes to locate tilted objects will cover many redundant areas (i.g., background or adjacent objects), especially in dense object scenes as shown in Fig.1(a). In this case, axis-aligned bounding boxes are not conducive to the processing of non-maximum suppression (NMS) and are prone to missing detection. Therefore, although many previous methods achieve the state-of-the-art in natural scene, they are not quite suitable for object detection in aerial images.