基于SE_ResNeXt-50的小麦不完善粒分类研究  

Research on Classification of Imperfect Grains of Wheat Based on SE_ResNeXt-50

作　　者：熊浩添 王鹏博 刘亚孰 蒋玉英[1] 王飞[2] 高辉[2] XIONG Haotian;WANG Pengbo;LIU Yashu

机构地区：[1]河南工业大学人工智能与大数据学院,郑州450001 [2]河南工业大学信息科学与工程学院,郑州450001

出　　处：《科技创新与应用》2023年第22期47-51,56,共6页Technology Innovation and Application

摘　　要：针对目前传统小麦不完善粒检测误差大、效率低、麦粒易受损坏等问题。提出一种基于分组卷积残差神经网络的小麦不完善粒分类模型。通过嵌入挤压激励模块(SE)改进ResNeXt-50分组卷积残差网络结构,将ResNeXt-50网络的残差输出结果进行挤压、激励后与原结果进行通道间相乘,提升网络对不同支路权重的感知。改进后模型可以更好地学习通道间的非线性相互作用和非互斥关系,提升模型训练效率,提高准确率。模型识别准确率可达96.12%,为小麦不完善粒的分类识别研究提供一种新的技术支持,进一步为国家储粮安全提供理论支持。In order to solve the problems of traditional imperfect grain wheat detection error,low efficiency,wheat grain easy to damage and so on.A imperfect grain wheat classification model based on grouped convolution residual neural network is proposed.The structure of the ResNeXt-50 packet convolution residual network is improved by embedding the squeeze excitation module(SE).The residual output of the ResNeXt-50 network is squeezed and multiplied with the original results to enhance the network's perception of different branch weights.The improved model can better learn the nonlinear interaction and non-mutual exclusion between channels,improve the training efficiency and accuracy of the model.The recognition accuracy of the model can reach 96.12%,which provides a new technical support for the classification and identification of imperfect grains of wheat,and further provides theoretical support for the safety of national grain storage.

关 键 词：小麦不完善粒 图像分类 ResNeXt-50 无损检测 图像处理技术 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]