基于改进U-Net的水稻叶片细胞分割方法研究  

Research on Improved U⁃Net Method for Rice Leaf Cell Segmentation

作　　者：韦钙兴 易文龙[1] 刘昱成 赵应丁[1] 陈庭倬 WEI Gaixing;YI Wenlong;LIU Yucheng;ZHAO Yingding;CHEN Tingzhuo(School of Software,Jiangxi Agricultural University,Nanchang 330045,China;Biotechnology Research Institute,Chengdu New Chaoyang Crop Science Co.,Ltd.,Chengdu 611630,China)

机构地区：[1]江西农业大学软件学院,江西南昌330045 [2]成都新朝阳作物科学股份有限公司生物技术研究院,四川成都611630

出　　处：《河南农业科学》2023年第3期153-160,共8页Journal of Henan Agricultural Sciences

基　　金：国家自然科学基金项目(61762048);江西省自然科学基金项目(20212BAB202015)。

摘　　要：针对水稻叶片细胞图像中存在边界模糊、信噪比低和叶肉细胞相互粘连堆叠等特点导致分割精度不高的问题,提出一种基于改进U-Net的水稻叶片细胞分割方法。首先,将BA模块引入到ResNeXt网络中组成BAResNeXt模块作为网络的编码器,在提取深层的语义特征时提高网络对叶肉细胞的关注度;其次,在编码器与解码器之间加入通道交叉注意力机制,缓和解码器和编码器的语义歧义来增强分割图像特征的信息融合;最后,在解码器上采样阶段中使用SE注意力机制,以便过滤分割图像背景的干扰信息。为了验证该方法的有效性,将其与U-Net、Res-UNet、U-Net++和Deeplabv3+等深度学习网络进行试验比对,结果表明,改进的方法在水稻叶片细胞分割中表现最好,其查准率为96.03%、召回率为97.67%、准确率为97.47%、交并比为93.96%,相似系数为96.78%,均比其他网络高。Aiming at the issues of blurred boundaries,low signal⁃to⁃noise ratio,mutual adhesion and stacking of mesophyll cells in the image of rice leaf cells,which lead to low segmentation accuracy,we propose an improved U⁃Net method for rice leaf cell segmentation.Firstly,the bridge attention(BA)module is introduced into the ResNeXt network to form the BAResNeXt module as the network encoder to improve the network’s attention to mesophyll cells when extracting deep semantic features;secondly,a channel cross⁃attention mechanism is added between the encoder and the decoder to ease the semantic ambiguity between the decoder and the encoder to enhance the information fusion of the segmented image features;finally,the SE attention mechanism is used in the upsampling phase of decoder to filter the interference information of image background.In order to verify the effectiveness of the method,it was compared with deep learning networks such as U⁃Net,Res⁃UNet,U⁃Net++and Deeplabv3+.The results showed that our method had the best performance in rice leaf cell segmentation.The Precision(96.03%),Recall(97.67%),Acc(97.47%),IoU(93.96%)and Dice(96.78%)of our method were all higher than other networks.

关 键 词：水稻叶片 细胞分割 U-Net 注意力机制 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] S126[自动化与计算机技术—计算机科学与技术]