Research And Application Of Garbage Classification Based On Convolution Neural Network

With the improvement of the consumption level of our residents,the amount of garbage produced is increasing day by day.If the garbage is not classified in time,it will pollute the environment and even affect social governance.Due to the variety of garbage,manual sorting is not only bad environment,but also time-consuming and subjective.Compared to the traditional methods,Convolutional neural network only needs image preprocessing,feature extraction and full connection layer operation to classify garbage images,the model training is simple and easy to optimize.but because of various kinds of don’t spam images with similar characteristics between local or data imbalance,causing the fine-grained image classification problem is ignored,which affect the classification accuracy.In addition,due to the limitation of target detection model,there are some problems in the detection of multi category garbage objects in the image,such as missed detection,false detection,low generalization,and lack of ealtime recognition application platform.In view of the above problems,this paper uses convolutional neural network to conduct research on garbage image classification and target detection.The main contents are as follows:1)Garbage image classification.Aiming at the problem of low classification accuracy of single waste type in the image,Xception model was selected to improve its structure and optimize related parameters.First,there is a reduction in accuracy caused by too small batchsize through Group Normalization to replace Batch Normalization between layers of the network model.Secondly,the se-Net module was added to the last part of Xception model to strengthen important channel information.Remove the original full connection layer,add the global average pool layer,deny layer and dropout layer in turn,reduce the number of parameters and improve generalization.Finally,parameters of the improved model were optimized,and Nadam optimizer and cosine annealing learning rate were used to update network parameters iteratively,and regularization methods such as label smoothing were used to improve generalization.Experimental results show that the proposed method can effectively improve the performance of garbage image classification through transfer learning.2)Garbage target detection.Aiming at the problem of low performance in detecting multiple kinds of garbage in the image,the traditional detection model SSD is selected to improve its structure and optimize relevant parameters.Firstly,the basic network VGG-16 in SSD was replaced with an improved Xception structure to reduce the number of parameters and improve feature extraction capability.Secondly,the feature fusion module is added in the detection layer to enhance the context information connection between the high and low layers and reduce the missing and error detection of small targets.Finally,Focal Loss,GIo U loss function and soft NMS are introduced to solve the sample imbalance problem.Experiments show that this method effectively improves the performance of garbage target detection.3)Garbage classification and detection platform.The Flask framework leverages HTML and CSS for page design and invokes optimized network models directly from the backend deployment.The platform mainly realizes the functions of uploading garbage image,returning classification and recognition results and popularizing garbage knowledge,providing users with a method to quickly identify garbage types.