Research On Deep Learning Based Harmful Garbage Detection And Classification Algorithm

With the improvement of human living standard,the types and forms of garbage are becoming more and more diverse and complicated.Among them,the harmful garbage is the most harmful and the most troublesome to deal with.However,the current research shows that the harmful garbage in different areas has only been classified and investigated,but there are few studies on the image of harmful garbage.At present,there are only large categories of garbage classification in the open garbage data sets,resulting in a small number and type of harmful garbage data sets.For the above problems,this paper will study the harmful garbage detection and classification algorithm based on deep learning.The main research contents are as follows:(1)Analyze the types of harmful wastes and divide them into 5 categories and 19 subcategories.The initial data set is obtained by network,and the image is preprocessed by geometric transformation(translation,rotation,etc.)and color transformation(sharpening,Gaussian noise,etc.).The images of the expanded dataset were normalized and numbered again,and then the annotation tools were analyzed and selected,and the data were annotated.(2)Hardware resource configuration and software environment construction,one-stage and twostage detection algorithms are theoretically analyzed,and YOLOv3 and YOLOv5 networks in one-stage are selected for comparative experiments.Experimental results show that YOLOv5 model is better than YOLOv3 model in speed and accuracy.YOLOv5 was selected for this detection classification.(3)Improve YOLOvS network.Firstly,the hybrid attention of global average pooling(GAP)and global maximum pooling(GMP)was used to improve its extraction ability and get better results.Then,target box weighted fusion(WBF)was used to replace non-maximal suppression(NMS),and the experimental results were compared.Through experiments,it is found that both YOLOv5s+Attention and YOLOv5s+WBF have improved accuracy,while YOLOv5s+Attention+WBF has the highest accuracy,which has improved accuracy by 2.11%compared with YOLOv5s.