Object detection methods for high resolution remote sensing images
The high resolution of satellite remote sensing images and the small relative size of the target within the image make it difficult to ensure both detection accuracy and operation speed.In order to solve the problem of target detection in high pixel remote sensing images,this paper proposes a detection method that combines sliding window segmentation and a small target detector.Firstly,the image is segmented into multiple subgraphs using the sliding window method,the sliding step is slightly smaller than the size of the window to make each subgraph have a certain overlap between them,and a larger segmentation window is used to reduce the number of subgraphs segmented.After that,the subgraphs are compressed and the compressed images are processed using a target detection algorithm to reduce the running time of the algorithm.Finally,the detection results are merged and a non-maximization suppression strategy is used to remove the targets that are repeatedly detected in the overlapping parts.In terms of detection algorithm,based on YOLOv8n,this paper uses SPD convolutional kernel and NWD to improve the network structure,and adjusts the feature pyramid structure to improve the algorithm's performance in detecting small targets,which enables the algorithm to adapt to compressed subgraphs at larger sizes in order to reduce the number of image segmentation and improve the detection speed.The experiment proves that on the vehicle detection dataset with an average image resolution of 4 000×4 000,the average accuracy of the method for target detection is 55.7%,and the average computation time per image is 47.5 ms.The accuracy is improved by 17%compared to YOLOv8n,15%compared to YOLOv5s and 7%compared to YOLOv6s.The operation efficiency of the proposed method meets the real-time requirement,which is capable of detecting targets in satellite remote sensing images in real time with higher accuracy.
国家科技期刊平台
NSTL
万方数据
