桃树缺磷症(Peach Phosphorus Deficiency,PPD)初期症状不明显、不同阶段症状差异大,而现有的基于计算机视觉的桃树病害识别模型,识别准确率不高、对不同品种识别泛化性差,为此,提出改进Faster R-CNN(Faster Region based Convolutional Neural Network)模型.首先,使用 RS(Rank & Sort)-Loss 函数代替 区域建议网络(Region Proposal Network,RPN)中的交叉熵函数;其次,使用Soft-NMS(Non-Maximum Suppression)算法代替原有的NMS算法;最后,使用ResNeXt101网络替换原来的特征提取网络,提高对PPD识别的准确率和泛化性,并在自建PPD数据集上进行检测试验.试验结果表明:改进后的Faster R-CNN网络在自建PPD数据集上对PPD的各类别平均检测准确率达92.28％、召回率达92.31％、识别准确率达92.28％,满足实际应用要求.
Research on peach phosphorus deficiency detection based on improved Faster R-CNN
The initial symptoms of peach phosphorus deficiency(PPD)are not obvious,and the symptoms differ greatly in different stages.However,the existing peach disease identification model based on computer vision has low identification accuracy and poor generalization of different varieties.Therefore,the Faster R-CNN(Faster Region based Convolutional Neural Network)model is improved.Firstly,the RS(Rank & Sort)-Loss function is used to replace the cross-entropy function in the Region Proposal Network(RPN).Secondly,the Soft-NMS(Non-Maximum Suppression)algorithm is used to replace the original NMS algorithm.Finally,ResNeXt101 network is used to replace the original feature extraction network to improve the accuracy and generalization of PPD recognition,and the detection test is carried out on the self-built PPD data set.The experimental results show that the improved Faster R-CNN network has an average detection accuracy of 92.28％,a recall rate of 92.31％and a recognition accuracy of 92.28％on the self-built PPD data set,which meets the practical application requirements.
NETL
NSTL
万方数据
维普
