基于卷积神经网络的垃圾图像分类模型研究应用

Research and Application of Garbage Image Classification Model based on Convolutional Neural Network

传统的垃圾分类方法往往借助于传感器完成垃圾识别分类,存在分类的准确率不高、模型复杂、缺乏高效操作性等问题。为解决这一问题,提出结合卷积神经网络的垃圾分类方法。使用具有高效特征提取性能的Inception-V3和ResNet50两种卷积神经网络对华为公开垃圾数据集Garbage Date进行训练,建立垃圾分类模型。实验表明,在训练集上的Inception-V3和ResNet50训练的准确率分别为89.9%和95.1%,交叉熵损失函数分别为1.463和1.363。使用可视化界面验证测试集中随机6类单张图片,ResNet50的准确率均高于Inception-V3。但ResNet50却不及Inception-V3稳定,Inception-V3准确率曲线更平滑。Inception-V3的收敛速度也比ResNet50快,消耗资源更少。

Traditional garbage classification methods of ten rely on sensors to complete garbage identification and classification,but there are problems such as low classification accuracy,complex models,and lack of efficient operability.In order to solve this problem,a garbage classification method combined with Convolutional Neural Networks(CNN) is proposed.Two convolutional neural networks,Inception-V3 and ResNet50,with high-efficiency feature extraction performance,were used to train Huawei’s public garbage data set Garbage Date,and a garbage classification model was established.Experiments show that the accuracy rates of Inception-V3 and ResNet50 training on the training set are 89.9%and 95.1%,respectively,and the cross entropy loss functions are 1.463 and 1.363,respectively.Using the visual interface to verify the six random single images in the test set,the accuracy of ResNet50 is higher than that of Inception-V3.But ResNet50 is not as stable as Inception-V3,and the accuracy curve of Inception-V3 is smoother.Inception-V3 converges faster than ResNet50 and consumes less resources.