Research On Remote Sensing Image Rotation Object Detection Algorithm Based On Improved YOLOv7

Objects in remote sensing images are often diverse and complex,making it difficult for traditional remote sensing image object detection techniques to meet current needs.With more and more achievements in artificial intelligence,the application of deep learning-based object detection techniques in remote sensing images has become more and more widespread.Different from other images,the object instances in remote sensing images exist at any angle,the small object instances are densely distributed in some scenes,and the images have the characteristics of high resolution.The general object detection network is prone to miss detection,false detection and other problems.At the same time,the network Receptive field is small,and the detection performance is poor.To address the above problems,this thesis proposes a rotating object detection algorithm for remote sensing images,and the main work is as follows:(1)A rotation object detection algorithm for remote sensing images is proposed to address the characteristics of objects having arbitrary angles.A circular smooth label is used for angle classification to solve a series of problems in calculating object rotation angles.The network input size is also optimized according to the characteristics of high resolution of remote sensing images,which improves the detection performance of the algorithm for small objects.Aiming at the problems of fuzzy regional Semantic information,insensitivity to feature details,and easy to cause object loss in the traditional method of up sampling the feature map with the nearest neighbor interpolation,a new up sampling method,namely,the depth separable deconvolution based up sampling method,is used.When expanding the feature map,it can diversify the expanded pixels,so that the obtained feature information is more abundant,Improve the accuracy of the network model without increasing algorithm complexity.Experiments on the DOTA dataset show that the proposed DC-YOLOv7 algorithm increases m AP by 3.3% compared to the original algorithm,with only an increase of 3.1 GFLOPs.(2)Aiming at the problems of insufficient use of local context information of remote sensing objects and small receptive field,a feature extraction method based on cavity convolution context transformation network is proposed.It can expand the receptive field while better extracting the feature information of objects in remote sensing images,and improve the detection performance of the network.Meanwhile,the correlation channel inthe compressed channel attention mechanism of the frequency channel attention network is utilized to increase the perception ability of the network model for the object feature information and better capture the detail information in the image,thus improving the accuracy and robustness of the whole algorithm.By conducting experiments on different remote sensing image datasets,the m AP reaches 78.35%,the FPS is 39.1 for the DOTA dataset,and the m AP reaches 94.1% for the Hrsc2016 dataset.This proves that the DFDCYOLOv7 algorithm proposed in this thesis has strong robustness and excellent detection performance for the detection of rotating objects in remote sensing images.Figure 28 Table 8 Reference 88...