Lightweight Research On Ground Small Target Detection Algorithm Of Low-altitude UAV

Low-altitude drones can save a lot of manpower and time costs in secur ity patrols and inspections due to their unique vie wing angle.Howe ver,long-distance aerial photography has small size and blurred appearance.The traditional target detection algorithm faces great challenges in the field of small target detection.With the collection of massive data in the information age,deep learning algorithm gradually replaces the traditional target detection algorithm and occupies an advantage,but it is difficult to deploy on the equipment with limited resources because of the huge amount of computation of its model.This paper mainly realizes the improvement and optimization of the small target detection algorithm for aerial photography on the premise of reducing the amount of model calculation.（1）This paper chooses the YOLOv3-Tiny network model as the backbone network,and uses the image feature pyramid structure to expand based on the backbone network to achieve multi-level feature fusion prediction.In order to eliminate the loss of spatial feature information hierarchy caused by the pooling operation and the multi-scale operation in the multi-layer structure in the convolution process,a spatial attention mechanism based on the residual structure is proposed.（2）Aiming at the problem of low boundar y regression accuracy in One-Stage target detection algorithm,this paper proposes a boundary regression model based on a reward mechanism.Introduce the idea of reinforcement learning to guide the bounding box regression based on the coarse positioning of the original bounding regression strategy,and use the variant Io U calculation method as the evaluation index of the reward function for fine adjustment.（3）This paper adopts the idea of deep separable convolution to lighten the network model.With the advantage of the unique lo w computational complexity of deep separable convolution,part of the standard convolution is replaced by separable convolution based o n the residual structure,thus reduce the amount of calculations and parameters of the convolutional layer,and obtain a lightweight small target detection model.This paper selects the visdrone2019 data set as the experimental data support.The experimental results show that the m AP value of the improved small target detection model in this paper is 33.15%,and the boundary regression accuracy is increased by 23.74%.Under the premise of ensuring that the small target detection accuracy loss does not exceed 5%,the model compression ratio reaches 25.37%.