Research On Object Detection In Remote Sensing Images Based On Improved YOLOv5

Remote sensing image object detection is a key technology of remote sensing digital image processing and has important research value in military early warning and civil security.However,remote sensing image has complex imaging environment and large difference of object characteristics,which bring many difficulties to the object detection task,and the detection accuracy is also reduced.Existing remote sensing image object detection models have complex structure,large computing storage,high dependence on hardware resources,and low detection speed.In order to improve the above problems,this thesis made the research from two aspects of detection accuracy and detection speed.In order to solve the problem of missing detection and false detection in remote sensing images caused by complex background interference and large difference in object scale,this thesis designed an object detection algorithm based on attention and multi-scale feature fusion.Firstly,multi-attention are fused in the convolution block of the feature extraction part of YOLOv5 to make the network focus on the object key region and suppress interference information.After that,adding a small object detection layer,which can fully obtain the scale features of the image.Finally,multi-path fusion of local feature and global feature is carried out to enhance the interaction between the network and the scale feature information.The experimental results show that compared with the YOLOv5 network,the m AP of the improved network increased by 4.4% in DIOR dataset and 3.8% in NWPU VHR-10 dataset,which can effectively improve the detection effect.In order to solve the problem of large consumption of computing resources by existing remote sensing image object detection models,this thesis designed a remote sensing image object detection algorithm based on ConvNeXt.The improve ConvNeXt network structure replaced the YOLOv5 backbone network,while the neck network used an improved lightweight module to reduce the computational redundancy caused by the convolutional stack and improve the reasoning speed.The experimental results show that compared with the YOLOv5 network,the improved network reduces the parameters by 2.4M,reduces the calculation amount by 23.6%,and increases the detection speed by 13f/s,which is convenient to be applied to the detection equipment with limited computing resources.In this thesis,two improved algorithms are proposed using the YOLOv5 basic network,and good experimental results are obtained.It shows that the improved algorithms have higher performance and occupies less resources,which have practical value.