| With developing of unmanned aerial vehicles and space technology,small remote sensing image target detection and recognition is playing one key role to get information from the ground surface.The information can be used in the military,security,environmental protection,and other areas.Small remote sensing image target detection has problems with small target feature extraction,randomness in angle,a heavy network structure,and insufficient detection accuracy.The thesis focuses on the small remote sensing image target detection algorithm by using the YOLOv5 general object recognition algorithm as a starting point.The goal is to find the better solutions to the problems listed above.A new data augmentation process called Albup is proposed to solve the problems of small target feature extraction and detection accuracy in remote sensing images.It combines horizontal rotation,random cropping,and random brightness adjustment methods during the data preprocessing stage.Compared to standard augmentation methods,the mean average precision(m AP)rises of 1.9%.The network structure is improved by making a detection head for small targets.In order to make better use of the information from the extracted features,a 160*160 convolutional layer is added to the original base.After that,the old FPN structure iss replaced with the new feature pyramid,Bi-FPN.This method is different from other feature fusion methods because it uses weights to make the information from features with different scales more evenly balanced.Multi-scale feature fusion is applied to make the recognition work better.The results of the experiments show that the optimized method improves the precision and m AP performance in the DOTA dataset by 1.4% and 0.8%,respectively,and improves the frames per second(FPS)performance by 2.9.In the end,the backbone network is replaced by the Efficient Formerv2 network,which uses a local attention mechanism to help maintain high performance and improve detection accuracy,reaching a m AP of 81.3%.A novel rotation target detection method is suggested in response to the arbitrary angles of small targets in remote sensing images,to overcome the detection issues experienced by traditional horizontal detection boxes in remote sensing images.The method is based on oriented bounding box(OBB)detection boxes and can be successfully applied to the YOLOv5 network.The Ghost Wasserstein distance(GWD)loss function is implemented to alleviate boundary issues met during the rotation target detection process.Unlike the traditional inductive regression loss function,the GWD Loss function offers a regression loss technique that leverages Wasserstein distance.This includes converting the rotation boundary box into a two-dimensional Gaussian distribution and eliminating the inconsistency between regression under boundary and non-boundary positions.The experiments show that when compared to the loss function before optimization,the application of the GWD Loss method improves the precision and m AP performance by 5.6%and 1.4%,respectively,in the DOTA dataset.Leveraging the advancements made in the previous sections,the M-YOLOv5 remote sensing image small target detection algorithm is proposed,which is lightweight and utilizes the Gsconv method to reduce computational costs.The algorithm initially achieved an accuracy of 83.6%? however,a reduction in accuracy by 0.2% resulted in a decrease in parameter volume from 6.45 M to 5.21 M,along with an FPS increase from 63.3 to 64.1.The M-YOLOv5 algorithm not only overcomes the challenge of extracting small target feature information from high-resolution images and addressing the problem of ignoring dense and small targets but also enhances the detection accuracy of remote sensing image small targets effectively. |
