基于改进ShuffleNet的板栗分级方法

Chestnut Grading Method Based on Improved ShuffleNet

传统的板栗分级方法主要依靠人工或机械的多级振动筛,不仅分级准确率低而且容易把坏的板栗分成好的板栗。针对传统板栗分级存在的问题,构建轻量级的卷积神经网络实现高精度的板栗的自动分级。在自然光条件下用小米Note9手机拍摄获取包含优等品、一等品、合格品、虫蛀品和霉烂品板栗的5481幅图像应用于卷积网络模型的训练、验证和测试。在学习ShuffleNet的基础上构建了一个浅层卷积神经网络Shnet-1,Shnet-1由2个卷积模块和4个Shuffle构成的板栗图像特征提取网络。特征提取网络连接板栗分类器,分类器由全局平均池化层、隐含层和输出层组成的多层感知器。为了实现板栗分类的最大精度和最小计算量,对Shnet-1模型的超参数进行了优化。将Shnet-1的分类性能与各种深度学习模型如AlexNet、Mnet-1、ResNet18进行了比较分析。浅层卷积神经网络Shnet-1网络模型应用于板栗分级的准确率达到98.90%,坏的板栗被分为好板栗的比例小于0.5%。Shnet-1的计算量小,板栗图像分类时间为26 ms,其权重仅占488KB的物理存储容量。改进ShuffleNet的卷积神经网络模型Shnet-1模型能够快速和准确地完成对板栗的分级,为板栗的自动化分级提供了智能决策支持。

The traditional chestnut grading method mainly relies on manual or mechanical multi-stage vibrating screens,which not only have low grading accuracy but also easily divide bad chestnuts into good ones.In response to the problems in chestnut grading,a lightweight convolutional neural network was constructed to achieve high-precision automatic grading of chestnuts.Under natural light conditions,a total of 5481 images of high-quality,first-class,qualified,worm-eaten,and moldy chestnuts were captured using a Xiaomi Note9 mobile phone and applied to the training,validation,and testing of convolutional network models.On the basis of learning ShuffleNet,a shallow convolutional neural network Shnet-1 was constructed.Shnet-1 is a chestnut image feature extraction network composed of two convolutional modules and four Shuffles.The feature extraction network connects the Chinese chestnut classifier.The classifier is a multilayer perceptron composed of a global average pooling layer,a hidden layer and an output layer.In order to achieve the maximum accuracy and minimum amount of calculation of chestnut classification,the hyperparameter of Shnet-1 model was optimized.The classification performance of Shnet-1 was compared and analyzed with various deep learning models such as AlexNet,Mnet-1,and ResNet18.The accuracy of applying the shallow convolutional neural network Shnet-1 network model to chestnut grading reaches 98.90%,and the proportion of bad chestnuts being classified as good chestnuts is less than 0.5%.The computational complexity of Shnet-1 is small,with a classification time of 26 ms for chestnut images,and its weight only accounts for 488KB of physical storage capacity.The improved ShuffleNet convolutional neural network model Shnet-1 can quickly and accurately grade chestnuts,providing intelligent decision support for automatic grading of chestnuts.