Research On Garbage Image Classification Based On Improved Convolution Neural Network

In recent years,the sharp rise in the amount of domestic waste has caused great influence on the environment,so it is particularly important to properly classify and treat it.The advent of artificial intelligence era promotes the development of deep learning.Convolutional neural network is favored by many scholars because of its excellent feature extraction ability.Therefore,this paper studies and improves the garbage classification model based on convolutional neural network.Firstly,build a garbage image data set.The data sources of this study are mainly the garbage classification challenge cup of Huawei cloud AI competition and the web crawler of Baidu Library,which contains 19023 pictures,including other garbage,recyclable garbage,kitchen waste and harmful garbage.In addition,in order to improve the robustness of the model,data enhancement technology is used to expand the data set,mainly including geometric transformation,color transformation and adding noise.Finally,the image data size and data distribution are normalized and standardized.Secondly,Res Net and Res Next series of networks are studied based on Py Torch framework,and the experimental comparison is carried out based on transfer learning and model fine-tuning.The model with high classification accuracy is selected as the basic network,and Res Next101＿32x16d WSL model is finally selected.In addition,experiments show that the learning speed of the model is faster and the effect is better when using transfer training.Then,to resnext101＿32x16d WSL model is improved.Firstly,Dropout algorithm is added in the full connection layer to solve the problem of model over fitting;Secondly,Group Normalization is used to replace the Batch Normalization method to solve the problem of unstable model accuracy when the batch size is small;Finally,CBAM attention mechanism is added to the model,and channel attention and spatial attention are used to make the model focus on target features more accurately.Finally,based on the experiment,the following conclusions are drawn:(1)using Dropout can effectively alleviate the problem of model over fitting,and the accuracy of its verification set is basically the same as that before adding dropout.For the data set of this study,the effect is the best when the rejection rate is 0.5;(2)When the batch size is small,the accuracy of the model using Group Normalization is high,and the loss value fluctuates less in the training process,and the effect of the model is better than Batch Normalization;(3)CBAM can also improve the performance of models with complex structure and many layers,but pay attention to the location of addition.Experiments show that adding attention mechanism in the first layer of the network will make the model miss more important features and reduce the accuracy of the model.Adding CBAM in the last layer of the network has the best effect,and the classification accuracy is improved by nearly 1.5%.