Optimization Of Intensive Object Detection Algorithm For High Resolution Color Remote Sensing Images

Due to rich semantic information,high frequency detail information and special imaging perspective,remote sensing image target detection has many problems,such as complex background information,dense array of target objects and random direction.Direct use of natural language image target detection algorithm will cause problems such as target identification error,inaccurate location,missing detection,misdetection and repeated detection.At present,the detection accuracy of vehicle targets applied to each algorithm in high-resolution remote sensing data set is less than 75%.In this paper,vehicle targets with different intensity in high-resolution remote sensing images are selected to optimize the YOLOv5 natural image detection algorithm,solve the problem of feature information processing and dense target distribution,and improve the detection accuracy.The research content mainly includes the following three aspects.(1)In order to improve the feature extraction capability of the backbone network,FEM,a feature enhancement module that can expand the sensing range of target features,is introduced.Firstly,the sensitivity field of single feature extraction is expanded without increasing the number of convolution parameters by using cavity convolutional networks with different voidality ratios.Secondly,depthseparable convolution is introduced to extract spatial information and fuse channels to reduce the computational amount of convolution.Finally,BN layer and SiLU activation function are introduced to prevent the overfitting of the training process and increase the network nonlinearity.According to the experimental study,compared with YOLOv5,when FEM module was introduced,the detection time of single image increased by 1.7ms/image on the basis of 9.6ms/image,and the accuracy increased by 5.3%on the basis of 62.5%.In order to ensure the fast detection speed,the precision measurement has been improved.(2)In order to improve the feature fusion ability of the neck network,Swin-Trans module with selfattention ability is introduced.Firstly,Swin-Transformer module builds global search through multi-head self-attention mechanism,which complements the efficient local search capability of CNN.Secondly,Swin-Transformer uses sliding Windows to divide the whole feature graph into blocks,and approximate the computational complexity to linear on the basis of guaranteeing the window boundary feature fusion.Finally,the Dropout layer is added to the Swin-Trans module to prevent training from overfitting.According to the experimental study,compared with YOLOv5 algorithm,after FEM module and SwinTrans module were introduced,the detection time of single image increased by 6.8ms/image on the basis of 9.6ms/image,and the accuracy increased by 8.8%on the basis of 62.5%.Among them,the detection accuracy of low density target,medium density target and high density target increased by 9.7%,10.3%and 6.4%respectively.(3)In order to solve the problem of low detection rate caused by random orientation of dense targets,a rotating target detection algorithm is designed.First,select the long side annotation method to mark the rotating target box.Secondly,the boundary problem caused by rotation Angle and the deficiency of the existing loss function which can alleviate the boundary problem are analyzed.Finally,a Loss function.R-ECIoU loss.which can alleviate boundary problems,was designed by introducing factors such as the size of the repeated area between the predicted box and the real box.the distance between the center point,the size of the frame shape and the rotation Angle.According to the experimental study,compared with FST-YOLOv5,after the introduction of rotating target frame,the detection time of single image is increased by 1.1ms/image on the basis of 16.4ms/image,and the accuracy is increased by 4.3%on the basis of 71.3%.Among them,the detection accuracy of low density targets,medium density targets and high density targets increased by 0.7%,2.9%and 9.5%,respectively.In this paper,from two aspects of improving the feature learning ability of the detection algorithm and designing the rotating target detection algorithm,the target feature expression ability and target positioning regression ability of the detection model are improved,and the precision of intensive target detection in high-resolution remote sensing images is improved.