基于Mask R-CNN农田杂草实例分割与叶龄识别方法  

Method for segmentation and leaf age recognition of farmland weeds based on Mask R-CNN

作　　者：权龙哲[1] 吴冰 毛首人 QUAN Longzhe;WU Bing;MAO Shouren(School of Engineering,Northeast Agricultural University,Harbin 150030,China)

机构地区：[1]东北农业大学工程学院,哈尔滨150030

出　　处：《东北农业大学学报》2021年第4期65-76,共12页Journal of Northeast Agricultural University

基　　金：国家自然科学基金项目(52075092);黑龙江省博士后科研启动金资助项目(LBH-Q19007)。

摘　　要：为在田间复杂环境中实现对杂草和玉米植株准确实例分割和叶龄识别获取,提出一种基于改进掩码区域卷积神经网络(Mask Regions with convolutional neural network features,Mask R-CNN)的植物叶龄获取方法。具体实施为构建包含不同天气(晴天、阴天、雨后)和不同采集角度(俯视、30°斜视、45°斜视)数据集,增强数据并用作网络输入。通过更换3个特征提取网络(ResNet-50、ResNet-101、MobileNetv2)、搭建多种不同尺寸区域建议框、非极大值抑制法(Non-maximum suppression,NMS)更换为Soft-NMS算法、RoIAlign代替RoI Pooling方法提高模型精度。测试田间复杂环境下杂草和玉米图像。结果表明,以ResNet-101为特征提取网络的改进深度学习模型具有良好分割性能和鲁棒性,阴天检测精度高于晴天和雨后,30°斜视检测效果优于45°斜视和俯视。分割模型AP50为0.730,高于现有DeepMask、MNC、Mask R-CNN分割模型精度,表明该方法可提高对杂草和玉米植株的实例分割和叶龄识别精度。In order to achieve accurate instance segmentation and leaf age identification of weeds and maize in a complex field environment, a plant leaf age segmentation method based on an improved Mask Regions with convolutional neural network features(Mask R-CNN) was proposed. The specific implementation was to construct a data set containing different weather(sunny, cloudy, rainy) and different acquisition angles(top view, 30° squint, 45° squint), and to enhance the data for network input.In order to improve the accuracy of the model, the following methods were used, such as replacing three feature extraction networks(ResNet-50, ResNet-101, MobileNetv2), building a variety of different sizes of regional suggestion boxes, replacing the Soft-NMS algorithm with a non-maximum suppression algorithm(Non-maximum suppression, NMS), and using RoIAlign instead of the traditional RoI Pooling.Tested weeds and maize images in a complex field environment, the results showed that the improved deep learning model with ResNet-101 as the feature extraction network had good segmentation performance and robustness, the detection accuracy of cloudy was higher than that of sunny and rainy,and the detection effect of 30° squint was better than 45° squint and top view angle. The AP50 of this segmentation model was 0.730, which was higher than the accuracy of the existing DeepMask, MNC,and Mask R-CNN segmentation models, indicating that this method could more accurately perform instance segmentation and leaf age recognition for weeds and maize plants.

关 键 词：叶龄 分割 Mask R-CNN 机器视觉 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] S451[自动化与计算机技术—计算机科学与技术]