少样本条件下的复杂叶片图像语义分割  

Complex leaf image semantic segmentation under condition of few samples

作　　者：汪庆 杜炜 马春 谷宗运 WANG Qing;DU Wei;MA Chun;GU Zong-yun(College of Medicine Information Engineering,Anhui University of Chinese Medicine,Hefei 230012,China)

机构地区：[1]安徽中医药大学医药信息工程学院,合肥230012

出　　处：《齐齐哈尔大学学报（自然科学版）》2022年第3期21-25,31,共6页Journal of Qiqihar University(Natural Science Edition)

基　　金：安徽省高校自然科学重点研究项目(KJ2020A0392);安徽中医药大学校级自然重点项目(2020zrzd16);安徽中医药大学校级自然一般项目(2020zryb09)。

摘　　要：针对少样本条件下复杂叶片分割精确度不高的问题,提出一种基于数据增强的图像语义分割方法。使用翻转、平移方法对训练集中的图像进行增强扩充,利用VGG19代替原SegNet语义分割模型的VGG16主干网络进行模型训练。实验结果表明,在包含180幅复杂背景叶片的图像数据集上,使用该方法的评价指数MPA和MIOU达到了98.02%和95.79%,相比未使用数据增强的原模型分别提高了9.96%和15.27%。To solve the problem that the segmentation accuracy of complex leaves is not high under the condition of few samples,a method of image semantic segmentation based on data enhancement is proposed.Use flip and translation methods to enhance and expand the images in the training set,and use VGG19 to replace the VGG16 backbone network of the original SegNet semantic segmentation model for model training.The experimental results show that using this method on the image data set containing 180 leaves with complex background,the evaluation indexes MPA and MIOU reach 98.02%and 95.79%,which are 9.96%and 15.27%higher than the original model without data enhancement.

关 键 词：语义分割 SegNet模型 VGG19 少样本 复杂叶片 数据增强 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]