Research on remote sensing object detection based on improved PP-YOLOE-R

With the development of remote sensing technology, we are to obtain a large number of multi-temporal, multi-dimensional, and high-resolution remote sensing images. Remote sensing image targets have the characteristics of multi-scale, multi rotation angle, and complex background, and traditional object detection methods are difficult to obtain ideal results. Deep learning techniques, especially deep convolutional neural networks, have been widely applied in the field of remote sensing image object detection due to their powerful feature learning and generalization abilities. They are used to adapt to object detection at different scales and angles, extract target features of interest, and are crucial for the analysis and application of remote sensing images. Faced with this challenging problem, this article has made improvements on the baseline of PP-YOLOE-R. Firstly, in order to improve detection speed, PPHGNetV2, which is more suitable for rotation invariance requirements, is adopted; In order to maintain the original feature output dimension, Omni Dimensional Dynamic Convolution, also known as ODConv, is used in the neck section to reduce information loss during down-sampling process and improve the model's feature extraction ability; The experiment was conducted on the public dataset DOTA-v1.0, and the improved algorithm achieved an accuracy of 81.89 mAP, showing an improvement of 4.42 mAP compared to the original baseline. The experimental results show that the remote sensing images corrected by the algorithm have more high-value targets and faster detection speed.