采用注意力机制与改进YOLOv5的光伏用地检测

Detecting land for photovoltaic development based on the attention mechanism and improved YOLOv5

针对光伏产业快速发展所产生的光伏用地检测与定位需求,提出了一种基于YOLOv5改进的光伏用地检测算法YOLOv5-pv。为实现复杂场景下光伏用地的快速精确检测与定位,首先在YOLOv5基础上引入加权双向特征金字塔以实现简单快速的多尺度特征融合从而强化对小目标的检测能力;其次引入Ghost卷积以保留冗余信息中有用的特征图信息;最后增加协同注意力机制提高算法对光伏用地的关注度以提高抗背景干扰能力。实验结果表明:YOLOv5-pv比YOLOv5召回率提高6.68百分点,平均精度提高4.43百分点。该方法对光伏用地检测效果较好,可为光伏用地检测研究提供新的实验参考。

In response to the detection and positioning demands for land for photovoltaic development due to the rapid growth of the photovoltaic industry,this study proposed a YOLOv5-pv algorithm for the detection of land for photovoltaic development based on the improved YOLOv5.For quick and accurate detection and positioning of land for photovoltaic development in complex scenes,the YOLOv5-pv algorithm adopted a weighted bi-directional feature pyramid based on YOLOv5 to achieve simple and fast multi-scale feature fusion,thereby enhancing the ability to detect small targets.Subsequently,the Ghost convolution was employed to retain valuable feature map information in redundant information.Finally,a co-attention mechanism was integrated to improve the algorithm's attention on the land for photovoltaic development,increasing its capacity to resist background interference.The experimental results demonstrate that YOLOv5-pv outperformed YOLOv5,with the recall rate and average accuracy improved by 6.68 percentage points and 4.43 percentage points,respectively.Therefore,the method proposed in this study can effectively detect the land for photovoltaic development,holding referential significance for relevant detection research.