Research And Implementation Of Garbage Detection And Classification System Based On Deep Learning

The increasingly prosperous industry,agriculture and commerce have created a large number of material resources,and greatly enriched people’s material life.However,the problem of waste production has become increasingly serious,which makes the popularization of waste classification imminent.With the progress of science and technology,it has been a general trend to use artificial intelligence technology to help people deal with various tasks.Among them,computer vision technology can help people classify garbage.Model volume is an extremely important index for transplanting and using deep learning algorithms.Therefore,this thesis proposes a lightweight garbage detection and classification network(Garbage Net,GBGNet)based on YOLOv4 target detection algorithm,which reduces the requirements for the storage capacity,power consumption and computing capacity of the device platform,and designs a mobile garbage classification system deployed in limited resources.The specific work is as follows:(1)Various lightweight backbone feature extraction networks are studied.Through experimental comparison,the lightweight Ghost Net is finally determined to be used as the feature extraction network.The Ghost Net-YOLOv4 target detection network is designed.The training method of migration learning is adopted to ensure the network’s feature extraction ability.(2)The Garbage Photo(GBGPhoto)garbage detection and classification data set was established.The garbage data set was obtained by means of web crawler,and was divided into four categories: hazardous garbage,dry garbage,wet garbage,and recyclable garbage.The garbage detection and classification data set was further expanded with data augmentation,and established with Label Img.The GBGPhoto garbage detection classification data was set.(3)A GBGNet for garbage detection and classification was constructed.The volume of the feature fusion network model is compressed to 1/3 of the original volume through the design of GM-CBL(Ghost Module-CBL)lightweight convolution block and the use of Ghost Net structure and depth separable convolution idea.With the coordinate attention mechanism used to improve the network detection accuracy,Focal Loss loss function used to alleviate the parameter imbalance problem,GBGNet for garbage detection and classification is obtained.(4)The garbage detection and classification system is designed and completed.The system can call the camera to take pictures of garbage and display its location information,garbage name and garbage classification in real time.The information can be transferred to the subsequent sorting device to complete garbage sorting.(5)An intelligent garbage classification system based on Android system was designed and completed,and the system was successfully deployed on the mobile terminal.The system detects and classifies garbage in pictures by taking photos,recording videos and reading albums,providing users with a tool to quickly identify garbage types.