Research On Drone-to-ground Target Detection Method Based On Lightweight Algorithms

With the development of drone technology,its flexible and efficient data collection ability has expanded the application scenarios of target detection.However,current target detection algorithms are mostly based on natural scene images,and in the field of drones,there are limitations in platform storage resources,small targets are more common,and targets have large scale variations,which results in poor performance of existing algorithms.In this paper,based on the limited storage and computing resources of drone platforms and the need for lightweight algorithm models,a drone-to-ground target detection method is proposed,focusing on the problems of feature loss of small-scale targets and large-scale variations across targets.Firstly,to address the problem of wasting computing resources caused by generating redundant features required to improve accuracy during feature extraction,this paper uses depth-wise separable convolution structures with lower computational complexity to replace traditional convolutions in residual structures while ensuring as much detection accuracy as possible,reducing the number of model parameters and making the model more lightweight.To address the issue of insufficient feature information extraction due to the reduced number of channels in the lightweight network,a frequency channel attention mechanism is introduced and a one-dimensional convolution is used to replace its fully connected layer.By enhancing key features,this compensates for the shortcomings of inadequate effective information extraction,thereby improving the model’s detection accuracy.Next,in response to the problem of classical object detection models often using concatenation channel dimension operations during multiscale feature fusion,ignoring the influence of different scale features during feature fusion,this paper introduces a multiscale weighted feature fusion method that highlights the importance of different-level features to further improve model performance.Furthermore,to address the problem of large parameter sizes of traditional object detection models that are not suitable for the limited storage resources of drone platforms,L1 regularization is used to optimize the sparse momentum pruning algorithm.The model is then compressed using an improved pruning algorithm,resulting in a high compression rate and emphasizing the storage advantages of the final model,making it more adaptable to drone platform storage and computing resources.Finally,this paper proposes a drone-to-ground target detection algorithm based on lightweight models,which enhances the accuracy of target detection while being as lightweight as possible,improving the accuracy of drone-to-ground target detection.Through comparative experiments,it is proven that compared to classical object detection models,the proposed algorithm not only performs better in target detection tasks but is also lighter than traditional models,making it better suited to the storage conditions of drone platforms.