The Research On Garbage Image Classification Based On Transfer Learning

Due to the late development of garbage classification in my country,people have great doubts about the classification of various kinds of garbage,which leads to insufficient enthusiasm for garbage classification and inaccurate classification.For the current garbage classification dilemma,this paper proposes to use image recognition technology to classify garbage images correctly,so as to help people complete garbage classification faster and more accurately.After analyzing the status quo of garbage classification and having an in-depth understanding of the current deep learning technology,this paper proposes a method of garbage image classification based on transfer learning,which assists people from the source of garbage classification by means of garbage image recognition.Solve the problem of misplaced garbage.However,to realize the effective recognition and classification of garbage images,there are still the following problems:1.There are fewer garbage image data sets due to open source.Using the Garbage Classify data set provided by Huawei Technologies,the data set contains more than 14,000 pictures.There are a wide variety of junk pictures into 40 categories and 4 categories.The sample data is not balanced,and there is a problem that the amount of data in different categories is about 8 times different.2.Although there are many spam images on the Internet,most of the data is not marked,and collecting marked data is a complicated,time-consuming and laborious process.3.Building and training a more complex neural network model requires a lot of computer resources.In view of the problems of the above garbage image classification tasks,combined with the current research on image classification problems,the application of transfer learning is proposed to solve the problem of garbage image classification,which mainly includes the following contents:1.Research the technical means and methods used in the field of image recognition and classification,learn the classic convolutional neural network structure model,analyze the advantages of several commonly used convolutional neural networks,so as to provide reference for network selection.2.The two strategies adopted by the pre-training model based on the transfer learning method are studied,and the solution method for the garbage classification based on the fine-tuning transfer learning is given.The transfer learning method is selected for the garbage image classification research,and the two different strategies of training the fully connected layer and the fine-tuning network are compared and analyzed,and the fine-tuning trainer is selected as the more suitable training method.3.Select the AlexNet,VGG16,and ResNet50 convolutional neural networks as the pre-training models for experimental analysis and comparison,select the ResNet50 network with higher accuracy to classify and recognize garbage images,change the fully connected layer to 1000 classifications into 4 classification problems,fixed shallow Layer network weights,using garbage calssify data set to fine-tune the network.Different loss functions,optimization functions and learning rates are selected for comparative analysis,and the optimal network is selected as the garbage classification network（Garbagenet）for classification and recognition,and finally a classification accuracy of 92%is obtained.4.At the end of the paper,suggestions are given for further in-depth and prospective research work,such as expanding the basic data set,adding a spatial attention mechanism,and adding a recognition feedback mechanism.