Garbage Classification Research Based On Transfer Learning

With the great improvement of people's lives,the problem of garbage disposal has become increasingly serious.The city produces a large amount of garbage in people's life and production every day.Most of the garbage is disposed of in a harmless way such as burying,burning,composting,etc.But there is still a considerable amount of garbage,especially domestic garbage is not treated scientifically and reasonably.It is discarded at will,causing different degrees of pollution to the surrounding environment such as soil,air and water resources.In order to effectively alleviate this problem,garbage classification is its key treatment link,which can significantly reduce the cost of garbage disposal,increase the value of garbage recycling,and make the best use of its contents.However,due to the variety of domestic garbage,many people are not sure about the exact type of garbage,which has caused some random behaviors and increased the cost of garbage disposal.In recent years,the convolutional neural networks technology has become closer to maturity.It hasbeen widely used in computer vision,especially image classification,which has achieved many results.Among them,the MobileNetV2 network model is a lightweight convolutional neural networks which designed for portable devices.The network uses a deep separable convolution structure,and introduces an inverted residual structure and a linear bottleneck unit,which greatly reduces the calculation amount and size of model training.Based on the above background,this paper proposed a garbage classification method based on transfer learning,and designs and develops mobile terminal garbage classification software based on transfer learning,aiming to assist in solving the problem of garbage placement errors from the source.In view of the characteristics of the large number of garbage image and the relatively insufficient number of samples,this paper uses the transfer learning method to classify garbage images based on the pre-trained model MobileNetV2 network.The pre-training model MobileNetV2 is fully trained on ImageNet which is a large dataset,so it has excellent feature extraction and classification capabilities.It transfers the feature maps extracted by the model on ImageNet data to a new target task,saving time in retraining the model.This model-based transfer learning method not only can solve the problem of insufficient samples,but also can improve the classification accuracy because more weight coefficients can be extracted from the pre-trained model.Finally,a mobile terminal garbage classification system based on transfer learning was developed by using the trained MobileNetV2 garbage classification model.And the model was embedded in Android software to assist and improve the accuracy of garbage classification in people's daily lives.Through in-depth study of transfer learning and convolutional neural networks theory and technology,this paper analyzes,designs,and develops a model transfer learning garbage image classification system.This system based on MobileNetV2 network and combined with the characteristics of garbage image data.It mainly includes the following contents:1.According to the research progress of transfer learning and convolutional neural networks,this paper combs out the engineering methods and technologies of transfer learning implemented on the convolutional neural networks model.2.The three strategies of the pre-training model based on the transfer learning method are studied,and a solution to the garbage classification based on the fine-tuning transfer learning is given.3.Comparison of image classification experiments for convolutional neural network models,including VGG16,ResNet50 and MobileNetV2.Then select the lightweight MobileNetV2 model which is suitable for portable devices through multiple angles of consideration.Through the pre-trained model on the ImageNet dataset to fine-tune the garbage classification task,a high classification accuracy is obtained.4.Developed mobile terminal garbage classification software based on transfer learning.Carried out detailed feasibility analysis and requirements analysis of mobile terminal garbage classification software,then load the trained MobileNetV2 garbage classification model into the Android platform,and a mobile terminal garbage classification software based on transfer learning is developed in combination with Java language.The software provides main functions such as taking pictures,and trash recognition and classification.