Research On Windmill Detection Based On Optical Remote Sensing Image

Wind energy is a clean and pollution-free renewable resource.The proportion of wind power in my country’s entire power structure is increasing day by day.It has become the third largest power source in my country and is growing rapidly.As of March 2022,the cumulative installed capacity of wind power in my country has reached 337 million kilowatts.How to accurately obtain the distribution of wind farms is of great significance for wind power investment monitoring and early warning,land occupation detection and energy clean consumption capacity evaluation,and building and assembling a large number of windmills is also one of the strategies for the national carbon neutral goal.Therefore,the planning and analysis of wind farms has become a research hotspot,and the important task that follows is the detection and research of windmills.As a sign of a wind farm,a windmill is a multi-scale target in high-resolution images.Due to the influence of image acquisition time,lighting conditions,and surface coverage,the characteristics of the target vary greatly,and it has a small area and a small number in the entire remote sensing image.And other characteristics,the detection is difficult,and the general ground object classification detection method is not suitable for the detection task of the windmill.Therefore,in response to the above problems,this paper improves the existing model and proposes a method suitable for windmill target detection in remote sensing images.The specific research contents are as follows:(1)The windmill data set of remote sensing images is constructed.The existing open remote sensing data set contains a small number of windmill images,and the windmill target in the images has a single feature and monotonous background,which cannot meet the training requirements of deep learning models.Therefore,this paper selects and downloads remote sensing images of several domestic and foreign municipal areas for clipping and labeling,constructs a windmill data set of remote sensing images with rich features and diverse backgrounds,and uses mainstream target detection algorithms to conduct experimental comparison of the constructed data set.YOLOv5,which has a good target detection effect on remote sensing windmill data set,is selected as the benchmark network model for this research.(2)An improved model based on multi-scale features of remote sensing images and small target detection is proposed.To solve the problem of large difference in target scale in remote sensing windmill images,PANet structure in YOLOv5 network is replaced by Bi FPN structure to strengthen the utilization rate of underlying features and improve the detection accuracy of multi-scale targets.In view of the difficulty in identifying small windmill targets in remote sensing images,a small target detection head based on Transformer is added to the detection module of YOLOv5 network to improve the recognition performance of small targets.In order to further improve the detection accuracy of densely arranged objects in remote sensing windmill images,the GAM attention mechanism was integrated into the neck network of YOLOv5 to enhance the network’s ability to extract global information from high-resolution images.(3)The object detection system of remote sensing windmill image is designed and developed.Based on the improved YOLOv5 model,a prototype windmill detection system is designed and implemented by using Django framework and vue framework.The model training,data management,and detailed database design were introduced,which verified the effectiveness of the remote sensing windmill target detection prototype system and had good application value.