Household Garbage Detection And System Design Based On YOLOv5s

In recent years,China has actively promoted the policy of garbage classification,but there are problems such as: people are lack of relevant knowledge and the high costs of assisting people in sorting garbage,etc.With the development of object detection,people can apply this technology to assist citizens with garbage sorting which provides a new way to promote the policy of garbage classification.At present,research on garbage classification faces some difficulties,such as a limited number of garbage categories,and poor performance in real-time detection and feature extraction of the model.To solve those problems,this thesis conducts a research on the structure and complexity of the model.Thus,we design and implement a garbage classification system based on CGPD-YOLOv5 s,which can detect multiple types of garbage.The main research works are as follows.(1)H2-Household-Garbage dataset which contains multiple garbage categories is constructed.First,we merge the "Huawei Cloud Cup" dataset and HGI30 dataset.Then,we eliminate some ambiguous and duplicate data.Next,we annotate the data in PASCAL VOC format.Finally,we expand the datasets by data augmentation.The H2-HouseholdGarbage dataset has 23518 images,contains 57 types of garbage,which can be divided into four categories: recyclable garbage,kitchen garbage,hazardous garbage,and other garbage.(2)A household garbage detection method based on the CGPD-YOLOv5 s models is proposed.In view of the problems of large size variation and overlapping occlusions in garbage dataset,and the complexity of the model.We propose the CG-YOLOv5 s based on Coordinate Attention(CA)mechanism and Ghost Net.The CA mechanism is introduced to obtain the global sensory field better and improve the feature extraction ability of the model.Moreover,Ghost Net is introduced to the lightweight design of the model.Specifically,Ghost Conv is used for feature extraction,and G-C3 module is constructed based on G-Bottle Neck to reduce the number of model parameters and improve inference speed.Since the improved CG-YOLOv5 s model still has redundant parameters and the detection accuracy has decreased,we propose the strategy of network pruning and knowledge distillation to optimize the model futher.The results of experiment on H2-Household-Garbage dataset show that the size of CGPD-YOLOv5 s is compressed by75.5%,the real-time inference speed is imporved by 46.8%,compared to YOLOv5 s.And the accuracy is close to YOLOv5 s.The above experimental results prove that CGPDYOLOv5 s has reduced the complexity of the model with less loss of detection accuracy,and is suitable for deployment on mobile devices.(3)A garbage classification system based on CGPD-YOLOv5 s is designed and implemented.First,we analyze the requirements of the system from functional and nonfunctional aspects.Then,we design the overall architecture,function modules,and database of the system.Finally,we implement the user subsystem and administrator subsystem of the garbage classification system by HTML,CSS,React,Antd,My SQL database,etc.Specifically,the user subsystem has implemented functions including registration and login,household garbage detection,text information retrieval,etc.The administrator subsystem has implemented functions including user information management,detection record management,feedback data audit,etc.(4)The garbage classification system is tested.We test the performance of the system from functional and non-functional aspects.The results indicate that all functional modules in system meet the design requirements.And Household garbage detection function is applied in designed system with good real-time performance and accuracy.Besides,the pages of the system are simple and beautiful,and the compatibility is good.