| The application of remote sensing image object detection technology is widespread in military,civilian,and other fields.With the continuous advancement of deep learning technology,remote sensing object detection methods based on deep learning have greatly improved their performance.YOLOv4 series algorithms are a universal deep learning object detection algorithm that is open and friendly,and performs well in detecting objects in normal scenes.However,in remote sensing image object detection tasks,due to the inherent characteristics of small object scales,dense distribution,and varied angles of remote sensing objects,the detection performance of YOLOv4 series algorithms for small remote sensing objects still needs to be improved.Therefore,based on the YOLOv4 series of algorithms,this paper is aimed at improving the problems of false detection,missed detection and model lightweighting in remote sensing image object detection.The main work of this paper is as follows:(1)Aim to improve the detection performance for small objects in remote sensing images,a high-precision detection algorithm DEBA-YOLOv4 is proposed in this paper.This algorithm improves three aspects of YOLOv4,including the backbone network,the output feature branches,and the attention mechanism,based on the detection characteristics of small objects in remote sensing images.Firstly,the dense module replaces the residual module in the original YOLOv4 backbone network,enhancing the ability to extract image detail features and shallow information of the backbone network.Secondly,a new large-scale prediction branch is added at the neck of the network to improve the detection accuracy of small objects in remote sensing images.Lastly,the EBAM module,a mixed-domain attention module,is designed to effectively integrate local and global information,better focus on small object information,and suppress irrelevant information interference.Experimental results show that DEBA-YOLOv4 achieves mean average precision(m AP)of 96.24% and 91.97% on the VHR-10 dataset and the DOTA dataset,respectively,improving by 5.09% and 3.83% compared to YOLOv4.(2)Aim to improve the detection accuracy while maintaining the complexity of neural network of the lightweight remote sensing object detection method,A lightweight object detection method EPA-YOLOv4-tiny is proposed in this paper.This algorithm is based on the lightweight network YOLOv4-tiny,which first incorporates the lightweight multi-scale attention mechanism PSA module into the Efficient Net feature extraction network,replacing the original YOLOv4-tiny backbone network,which enables the network to focus more effectively on the object area while lightweight,enhancing the feature expression ability.In the neck structure of YOLOv4-tiny,an adaptive feature fusion mechanism ASFF is added to adaptively learn small object features and enhance the fusion and utilization of various scale features.The K-means++ clustering algorithm is used to optimize the prior box size that is more suitable for remote sensing image object detection.Experimental results on the VHR-10 and DOTA datasets show that EPAYOLOv4-tiny improves the mean average precision(m AP)compared to the original YOLOv4-tiny algorithm by 8.24% and 6.32%,respectively.Compared with the original YOLOv4 algorithm,the detection speed FPS is increased by 25frame/s and 20frame/s,respectively,and the parameter quantity is reduced by 37.01 M.This method maintains the lightweight scale of the YOLOv4-tiny model,but has higher detection accuracy and is a balanced algorithm of accuracy and speed.Figure [40] table [9] reference [65]... |
