AUTHOR=Pei Wenjing , Shi Zhanhao , Gong Kai 

TITLE=Small target detection with remote sensing images based on an improved YOLOv5 algorithm

JOURNAL=Frontiers in Neurorobotics

VOLUME=Volume 16 - 2022

YEAR=2023

URL=https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2022.1074862

DOI=10.3389/fnbot.2022.1074862

ISSN=1662-5218

ABSTRACT=Currently, small target detection in remote sensing images is a challenging topic. However, in remote sensing images, the target size may be small and the backgrounds may be complex and fuzzy. Thus, in this study, the YOLOv5s algorithm is modified and a new approach is proposed. First, to overcome the above challenges, the LCB module, which is composed of the LSM and C3 modules, is employed as the feature extraction module. Moreover, the SPPS and Dres2 modules are introduced. Multi-scale feature fusion can be realized and the performance of small target detection can be enhanced. Additionally, the input size of the network is increased from 640×640 to 1024×1024, and the output feature map size is set as 32×32, 64×64, and 128×128. Finally, the EIoU loss is used as the loss function to increase the convergence speed. The DIOR-VAS and Visdrone2019 datasets are selected for the experiments. The data enhancement strategies are applied to expand the remote sensing image training sets and to diversify the data sets. The ablation experiments are designed to verify the effectiveness of the proposed algorithm. Moreover, the proposed method is compared with the other five methods, and the effectiveness of small target detection is verified and comprehensively analyzed.