Research On Target Detection Methods For Aerial Images Based On YOLOv5

Target detection based on aerial images has important practical significance and can be applied in many fields such as ecological monitoring,traffic planning,natural disaster monitoring and rescue,military reconnaissance,etc.In recent years,the emergence of deep learning technology has brought more possibilities for detection in the field of aerial images,which can achieve fast image processing and high-precision target detection.The YOLOv5 algorithm performs well in target detection tasks,and has the advantages of high efficiency,accuracy,versatility,and high degree of technical integration.Therefore,this paper proposes a series of methods to improve the accuracy of aerial image target detection based on the YOLOv5 algorithm.Specifically,it can be summarized as the following three tasks :(1)This paper proposes a YOLOv5-based method for detecting rotating targets in aerial images,addressing the issue of arbitrary target orientation.Traditional YOLOv5 algorithms often encounter the problem of mismatched target labels and oversized aerial image pixels when directly detecting aerial images.To resolve these issues,this study first segments aerial images into uniform-sized images and transforms image labels using the long-edge definition method.The transformed label format of aerial images includes an angle label in addition to the label format of natural images.The processing of angle labels is added to both the pre-data processing and post-processing methods of the YOLOv5 algorithm.However,predicting the angle label using regression has bound problems.To solve this,the study uses the circular smoothing label method to process the angle label and converts the angle label using a Gaussian function,enabling the conversion of a regression loss to a classification loss.Experimental results show that the YOLOv5 algorithm based on aerial image detection outperforms other rotating target detection algorithms on the DOTAv2.0image dataset.Combining circular smooth labels with YOLOv5 effectively improves the model’s detection accuracy on DOTAv2.0.(2)This paper proposes the Trans＿YOLOv5 algorithm for detecting dense small targets,complex target backgrounds,and large differences in target scales in aerial images,achieved by improving the network structure of the algorithm.This study analyzes three types of difficulties in aerial imagery and proposes corresponding solutions.Specifically,the network structure of YOLOv5 is improved by replacing part of the C3 module with the Swin Transformer module to better capture the interaction information between global semantic information and associated targets,enhancing the ability to detect small targets.Additionally,an attention module is added to enhance the feature representation ability and improve the robustness of the model in dealing with the complex target background.A new layer of prediction is added to address the problem of large differences in target scales.Finally,the Trans＿YOLOv5 algorithm is proposed by combining the improved version of YOLOv5 mentioned above.Comparative experiments on the DOTAv2.0 image test set show that the Trans＿YOLOv5 algorithm has advantages in solving various problems in aerial images,and the overall detection accuracy of aerial image detection using the Trans＿YOLOv5 algorithm has been improved.Therefore,the Trans＿YOLOv5 algorithm proposed in this paper effectively solves the difficulties in aerial image detection and has practical value.(3)To optimize network parameters and improve algorithm robustness,an appropriate loss function is crucial in the process of aerial image target detection.This article analyzes and studies the classification and regression loss functions of the Trans＿YOLOv5 algorithm in each process of the target detection process.The regression loss function is discussed in detail,focusing on the roles of the Io U Loss,GIo U Loss,DIo U Loss,CIo U Loss,and SIo U Loss functions.These loss functions are used to calculate the training loss of the network,and backpropagation optimizes the training results.In the classification loss function,the paper primarily studies the cross-entropy loss function,Focal Loss,and Equalization Loss.The cross-entropy loss function optimizes classification tasks,while Focal Loss and Equalization Loss address the problem of unbalanced categories and targets in the dataset.Through numerous comparative experiments on the DOTAv2.0 image dataset,it was found that using the CIo U Loss and cross-entropy loss function in the Trans＿YOLOv5 algorithm results in better detection performance.This paper proposes the YOLOv5 algorithm for rotating target detection.We then propose the Trans＿YOLOv5 algorithm by improving the YOLOv5 network structure.Finally,we analyze and study various loss functions of the Trans＿YOLOv5 algorithm.To demonstrate the effectiveness of these methods,we conducted several comparative experiments.