基于改进Faster-RCNN模型的无人机影像白喉乌头物种的检测

Detection of Aconitum leucostomum Species in UAV Images Based on Improved Faster-RCNN Model

伊犁地区的白喉乌头(Aconitum leucostomum)是危害草原生态和畜牧业安全最为严重的毒害草物种,为了实现精准快速检测白喉乌头,本研究用无人机航拍获取白喉乌头影像数据集,采用深度学习技术,在Faster-RCNN算法基础上,以VGG16为主干网络,根据领域知识和数据集特点优化锚框大小等超参数改善算法性能。通过优化Faster-RCNN的锚框大小和微调学习率后,模型在测试集上取得的平均准确率为67.24%,相对于基准模型提高了21.57%。

Aconitum leucostomum in Yili area was the poisonous grass which most seriously endangeris grassland ecology and animal husbandry safety.To realize accurate and rapid detection of Aconitum leucostomum,this study aims to acquire the imaging data set of Aconitum leucostomum by UAV aerial photography.Based on the Faster-RCNN algorithm and with VGG16 as the backbone network,the hyper-parameters such as anchor box size were optimized according to domain knowledge and data set characteristics using deep learning technology,so as to improve the algorithm performance.After optimizing the anchor box size of Faster-RCNN and fine tuning the learning rate,the average accuracy of the model on the test set was obtained,which was 67.24%,21.57%higher compared with the benchmark model.