垃圾分类数据的机器学习方法研究

【作者】 王洋；

【导师】 周亚晶；

【作者基本信息】 黑龙江大学 ， 应用统计， 2020， 硕士

【摘要】 自2019年起,我国新的生活垃圾管理条例正式开始施行,我们以后在扔垃圾前都要先将垃圾仔细分类,但是由于日常生活垃圾品类繁多,各种垃圾分类标准也是五花八门,一时间人们恐怕难以快速正确地进行垃圾分类。机器学习算法发展快速且在许多领域都有应用场景,利用算法对图像识别可以大大提高垃圾分类的效率,为人们的生活带来便利。本文利用网上搜集的垃圾图像数据集,首先以支持向量机和K近邻为代表的传统机器学习算法为基础,人工提取图像的灰度直方图、HOG算子、LBP算子数据并用上述两个算法对图像分类,两个算法的最高分类准确度分别达到了79%和85.1%,接着以卷积神经网络深度学习算法为例,在卷积神经网络的经典模型VGGNet模型的基础上创新地简化该模型,通过减少卷积核的数量、修改卷积核尺寸和卷积层数使得在模型分类准确度损失较低的情况下大幅度减少模型运行内存的花销,较大地提升了模型的训练效率。简化后的Mini-VGG模型分类准确度达到了93%,具备实用价值。最终得出结论:如果特征选取合适,传统机器学习算法在图像分类领域中的分类准确度不弱于深度学习。未来图像分类领域深度学习仍占主导地位,但传统机器学习算法仍将占有一席之地。更多还原

【Abstract】 Since 2019,China’s new living garbage management regulations have been implemented.After that we must carefully classify the garbage before throwing it away.However,because the variety of living garbage and various garbage classification standards,it may be difficult for people to quickly and correctly classify the garbage.With the rapid development of artificial intelligence,machine learning combined with image recognition technology has been widely used in all aspects of life.Using algorithms to realized images can greatly improve the efficiency of garbage classification and bring convenience to people’s life.This paper uses the garbage image data sets collecting from the Internet.At first,extracting GH descriptor,HOG descriptor and LBP descriptor of images based on traditional machine learning represented by support vector machine(SVM)and K-nearest neighbor(KNN),and uses these two algorithms to classify images,the highest accuracy has reached 79% and 85.1% respectively.Then,taking the convolutional neural network as an example,based on a classic model named VGGNet,this paper innovatively simplified VGGNet by reduce the number of convolution kernels and convolution layers,modify the size of convolution kernels to ensure that the model classification accuracy loss is low and the cost of running memory of the model is greatly reduced,which greatly improves the training efficiency of the model.the accuracy of Mini-VGG has reached 93%,so this method is valuable.It is can be concluded that the classification accuracy of traditional machine learning algorithm is basically the same as deep learning if the feature selection is appropriately.In the future,the field of image classification will still be dominated by deep learning,but the traditional machine learning algorithm will still has a place.更多还原