基于改进Mask-RCNN的桃树穿孔病检测研究  

Study on Peach Shot-hole Disease Detection Based on Improved Mask-RCNN

作　　者：胡彦军 张平川[1] 张彩虹[1,2] 陈旭 李珊[1] 杨莹 马泽泽[1,2] HU Yan-jun;ZHANG Ping-chuan;ZHANG Cai-hong;CHEN Xu;LI Shan;YANG Ying;MA Ze-ze(School of Computer Science and Technology,Henan Institute of Science and Technology,Xinxiang Henan 453003,China;School of Information Engineering,Zhengzhou Electric Power Technology College,Zhengzhou 451450,China)

机构地区：[1]河南科技学院计算机科学与技术学院,河南新乡453003 [2]郑州电力职业技术学院信息工程学院,郑州451450

出　　处：《沈阳农业大学学报》2023年第6期702-711,共10页Journal of Shenyang Agricultural University

基　　金：河南省科技厅科技攻关项目(222102210116)。

摘　　要：穿孔病是桃树常见的病害,分为细菌性穿孔病(bacterial_shot-hole_disease,BSD)和真菌性穿孔病(fungal_shot-hole_disease,FSD)。大多数果农凭借经验难以准确识别两种病害,因而贻误防治,造成减产。为解决这一难题,提出了基于Mask RCNN(mask region based convolutional neural network)的桃树穿孔病检测方法。该方法对Mask RCNN模型进行了三方面的改进:首先,将Sim-AM(simple,parameter-free attention module)机制融入到残差网络的每一层,使用能量函数对神经元分配三维权重,增强对穿孔病关键特征的提取能力;其次,对RPN网络重复计算识别框进行简化处理,从9种Anchor Box降为3种,提升计算效率;再次,用软性非极大值抑制算法(Soft-NMS,soft non-maximum suppression)替换NMS算法进行Anchor Box选取,提高对遮挡病斑的检测效果。该研究使用二次迁移学习法对模型进行训练,首先利用Kaggle平台上公开苹果叶穿孔病数据集进行训练学习,然后利用两个数据集具有相似特征空间这一特点,在自建桃树穿孔病数据集上进行二次迁移学习,提高了检测准确率。实验结果显示,改进后的Mask-RCNN模型对桃树穿孔病全部类别平均检测精度mAP达到94.1%,召回率达到93.5%。Perforation disease is a common disease of peach trees,which is divided into bacterial shot-hole disease(BSD)and fungal shot-hole disease(FSD).Most fruit growers have difficulty in accurately identifying them through their experience,which leads to mistaken prevention and control,resulting in yield reduction.To solve this problem,a peach tree perforation disease detection method based on Mask RCNN(mask region based convolutional neural network)was proposed.The method improved the Mask RCNN model in three aspects:firstly,the Sim-AM(simple,parameter-free attention module)mechanism was integrated into each layer of the residual network,and the energy function was used to assign three-dimensional weights to the neurons,which enhanced the extraction of key features of the perforation disease;secondly,for RPN network Repeated computation of recognition boxes,a simplification process was carried out to reduce the number of Anchor Boxes from nine to three,which improved the computational efficiency;again,the NMS algorithm was replaced with the Soft Non-maximum Suppression algorithm(Soft-NMS,soft non-maximum suppression)for the Anchor Box selection,which improved the detection of occluded diseased spots.The study used the secondary migration learning method to train the model,firstly using the publicly available apple leaf perforation disease dataset on the Kaggle platform for training and learning,and then using the feature that the two datasets have a similar feature space,the secondary migration learning was carried out on the self-constructed peach tree perforation disease dataset,which improved the detection accuracy.The experimental results showed that the improved Mask-RCNN model achieved an average detection accuracy mAP of 94.1%and a recall rate of 93.5%for all categories of peach tree perforation disease.

关 键 词：桃树穿孔病 Msak RCNN Sim-AM 迁移学习 

分 类 号：S433.4[农业科学—农业昆虫与害虫防治]