Pei Wenjing, Shi Zhanhao, Gong Kai
The Seventh Research Division and the Center for Information and Control, School of Automation Science and Electrical Engineering, Beihang University (BUAA), Beijing, China.
School of Information Science and Engineering, Shandong Agriculture and Engineering University, Jinan, China.
Front Neurorobot. 2023 Feb 8;16:1074862. doi: 10.3389/fnbot.2022.1074862. eCollection 2022.
Small target detection with remote sensing images is a challenging topic due to the small size of the targets, complex, and fuzzy backgrounds.
In this study, a new detection algorithm is proposed based on the YOLOv5s algorithm for small target detection. The data enhancement strategy based on the mosaic operation is applied to expand the remote image training sets so as to diversify the datasets. First, the lightweight and stable feature extraction module (LSM) and C3 modules are combined to form the feature extraction module, called as LCB module, to extract more features in the remote sensing images. Multi-scale feature fusion is realized based on the Res 2 unit, Dres 2, and Spatial Pyramid Pooling Small (SPPS) models, so that the receptive field can be increased to obtain more multi-scale global information based on Dres2 and retain the obtained feature information of the small targets accordingly. Furthermore, the input size and output size of the network are increased and set in different scales considering the relatively less target features in the remote images. Besides, the Efficient Intersection over Union (EIoU) loss is used as the loss function to increase the training convergence velocity of the model and improve the accurate regression of the model.
The DIOR-VAS and Visdrone2019 datasets are selected in the experiments, while the ablation and comparison experiments are performed with five popular target detection algorithms to verify the effectiveness of the proposed small target detection method.
由于目标尺寸小、背景复杂且模糊，利用遥感图像进行小目标检测是一个具有挑战性的课题。
在本研究中，提出了一种基于YOLOv5s算法的小目标检测新算法。应用基于马赛克操作的数据增强策略来扩展遥感图像训练集，以使数据集多样化。首先，将轻量级且稳定的特征提取模块（LSM）和C3模块相结合，形成称为LCB模块的特征提取模块，以提取遥感图像中的更多特征。基于Res 2单元、Dres 2和空间金字塔池化小型（SPPS）模型实现多尺度特征融合，从而增加感受野，基于Dres2获得更多多尺度全局信息，并相应地保留所获得的小目标特征信息。此外，考虑到遥感图像中相对较少的目标特征，增加并设置网络的输入大小和输出大小为不同尺度。此外，使用高效交并比（EIoU）损失作为损失函数，以提高模型的训练收敛速度并改善模型的精确回归。
实验中选择了DIOR-VAS和Visdrone2019数据集，同时与五种流行的目标检测算法进行了消融和对比实验，以验证所提出的小目标检测方法的有效性。
