Research On UAV Aerial Image Object Detection Algorithm Based On Deep Learning

With the progress of society and technology,the effect of enterprises’ reforming operation mode and improving operation efficiency with the help of drones has become more and more obvious,and "UAV+industrial application" has become a major development trend.There is a huge market demand for all kinds of applications related to unmanned aerial vehicles and their products,which brings space and motivation for the development of various technologies combined with unmanned aerial vehicles.In UAV intelligent system,it is an important task to combine computer vision technology based on deep learning with UAV aerial photography technology to realize target detection,and target identification and location is one of its most important links.Different from general target detection,UAV aerial images have the characteristics of high resolution,many and dense small-scale targets,large scale difference between different types of targets and complex background,so compared with other images,UAV aerial images make target detection more difficult and complicated.Aiming at the above problems,this paper designs a lightweight target detection algorithm based on improved YOLOv5,which is trained and verified on the public data set VisDrone-2019.The main research results are as follows:(1)In view of the small size and dense distribution of ground targets photographed by UAV at high altitude,based on the three detectors of the original YOLOv5 model,this paper adds a detector P2 specifically for small-scale targets,which allows the shallow high-resolution feature map to participate in the subsequent multi-scale feature fusion module,effectively avoiding the problem that the network loses the key feature information of small-scale targets in the process of multiple downsampling,and improving the detection accuracy of the model for small targets.(2)In this paper,an enhanced multi-scale feature fusion pyramid network DSI-FPN is designed.The FPN+PAN network is optimized by using depth separable Involution and involvement operators with less parameters and computation,and the spatial attention mechanism,so as to generate more informative feature maps for network detection tasks.Secondly,this paper proposes an adaptive channel spatial attention mechanism SCBAM,which introduces a self-attention mechanism into the CBAM module,and adds non-local information to the interaction that originally had only local information,breaking the limitation of convolution kernel,expanding the receptive field of the model and improving the feature expression ability of the model.(3)In order to solve the problem of insufficient computing power when deploying the target detector for UAV equipment,in the lightweight design part of the model,this paper draws lessons from the idea of knowledge distillation,and uses the framework of joint teacher network knowledge distillation based on feature layer.Design the distillation loss of joint teachers,balance the contribution of two teachers’ networks to the truth value,and dynamically adjust the learning trend of student networks.This mechanism excavates the potential knowledge in the network characteristic layer of teachers,and guides students to learn online with more universal characteristics.By transferring the knowledge information of the middle feature layer and the output layer of the joint teacher network,students can learn online,improve the detection accuracy,and effectively reduce the parameters and model size of the network.Finally,through experimental verification,the detection accuracy of the improved model based on YOLOv5m on the public UAV image dataset VisDrone-2019 reaches 43.9%,which is 7.4 percentage points higher than the original model.The number of parameters of the model after knowledge distillation was reduced by about 58 percentage points compared with the previous model,and the detection accuracy was 40.2%,which was 7.8 percentage points higher than that of the original YOLOv5s model.