基于深度学习的粘连米粒实例分割算法研究  

Research on Instance Segmentation Algorithm of Adhesive Rice Grains Based on Deep Learning

作　　者：尚玉婷 王粤[1] 刘彬 Shang Yuting;Wang Yue;Liu Bin(College of Information and Electronic Engineering,Zhejiang University of Industry and Commerce,Hangzhou 310018)

机构地区：[1]浙江工商大学信息与电子工程学院,杭州310018

出　　处：《中国粮油学报》2022年第9期224-230,共7页Journal of the Chinese Cereals and Oils Association

摘　　要：基于机器视觉的农作物外观质量检测近些年越来越受到人们的关注。在抽样检测台上由于米粒可能相互触碰、粘连,采集的图像若不进行分割预处理会造成后续大米外观品质评测的失误。由此,本实验提出了一种基于Mask R-CNN改进的实例分割网络RiceInstNet,用于粘连米粒的图像分割。主干网络由2个改进的VoVNetV2网络并行组成,在大规模减少网络参数的同时加强对粘连米粒图像的特征提取,另外在掩膜分支上增加了一个学习物体边界的子网络,利用边界特征丰富掩膜特征并促进掩膜预测得更加精细。实验结果表明,该网络可以实现对粘连米粒的实例分割,与Mask R-CNN网络相比,改进后的网络模型RiceInstNet的平均精度和召回率分别由87.2%、89.7%提升到90.4%、93.7%,同时本网络模型更轻量,非常适合集成到移动终端或嵌入式设备中。Crop appearance quality detection based on machine vision has attracted more and more attention in recent years.On the sampling inspection table,because rice grains may touch and adhere to each other,if the collected images are not segmented and preprocessed,it will cause errors in the subsequent evaluation of rice appearance quality.Therefore,this paper proposed an improved instance segmentation network named RiceInstNet based on Mask R-CNN for image segmentation of adhesive rice grains.The backbone network was composed of two improved VoVNetV2 in parallel,which not only reduced the network parameters on a large scale,but also strengthened the feature extraction of the adhesive rice grain image.In addition,a sub-network for learning the object boundary was added on the mask branch,which used the boundary features to enrich the mask features and promoted the mask prediction to be more precise.The experimental results showed that the RiceInstNet could achieve the instance segmentation of adhesive rice grains.Compared with Mask R-CNN,the average precision and recall of the RiceInstNet were increased from 87.2%and 89.7%to 90.4%and 93.7%respectively.Furthermore,the RiceInstNet was more lightweight and very suitable for integration into mobile terminals or embedded devices.

关 键 词：实例分割 轻量级 粘连 边界子网络 深度学习 

分 类 号：TP29[自动化与计算机技术—检测技术与自动化装置]