基于改进Faster-RCNN网络的无人机遥感影像桃树检测

Detection of peach trees in unmanned aerial vehicle (UAV) images based on improved Faster-RCNN network

研究桃树精准检测定位方法可为桃园精准化管理提供支撑,该研究利用无人机遥感结合深度学习算法对无叶期桃树进行检测,并对萌芽期桃树和花期桃树进行区分。在Faster-RCNN原始网络的基础上提出了3种改进:替换主干网络为融合卷积注意力模块(convolutional block attention module, CBAM)后的ResNeXt-50、感兴趣区域特征提取方法使用ROI Align代替ROI Pooling、引入不平衡交叉熵损失函数Focal Loss,并采用消融试验对这些改进方法的效果进行分析。实验表明,与未改进的Faster-RCNN相比,改进后的模型mAP(mean average precision)提升了8.95百分点,达到了86.46%,能够较好地区分花期桃树与萌芽期桃树。对模型提升贡献最大的改进是ResNeXt-50-CBAM主干网络替换,mAP提升5.98百分点;ROI Align的使用减少了特征量化过程中的误差,mAP提升2.33百分点。与其他主流检测模型YOLOv3、 YOLOv5x和SSD相比,该研究提出的模型检测效果更优。该研究提出的无人机遥感结合改进Faster-RCNN算法能够较好地对桃树进行检测,可以满足桃园精准化管理需求。

Studying precise detection and positioning methods for peach trees can provide support for precision management of peach orchards.This study utilizes unmanned aerial vehicle(UAV)remote sensing combined with deep learning algorithms to detect leafless peach trees and differentiate between budding and flowering peach trees.Three improvements are proposed based on the Faster R-CNN original network:replacing the backbone network with a ResNeXt-50 integrated with a convolutional block attention module(CBAM),using ROI Align instead of ROI Pooling for feature extraction from regions of interest,and introducing the Focal Loss function for imbalanced cross-entropy loss.Ablation experiments are conducted to analyze the effectiveness of these improvements.The experiments demonstrate that compared with the unimproved Faster R-CNN,the improved model achieves a mean average precision(mAP)increase of 8.95 percentage points,reaching 86.46%,enabling better differentiation between flowering and budding peach trees.The most significant improvement contributing to the model s enhancement is the replacement of the ResNeXt-50 backbone network with CBAM,resulting in a 5.98 percentage points increase in mAP;the use of ROI Align reduces errors in the feature quantization process,leading to a 2.33 percentage points increase in mAP.Compared with other mainstream detection models such as YOLOv3,YOLOv5x,and SSD,the proposed model in this study demonstrates superior detection performance.The UAV remote sensing combined with the improved Faster R-CNN algorithm proposed in this study can effectively detect peach trees,meeting the requirements of precision management in peach orchards.