Research And Implementation Of Garbage Classification And Recognition Methods Based On Deep Learning

With the continuous development of economy and technology,people's material needs are also increasing,along with the growth of people's garbage.Garbage problem has gradually entered the public's field of vision,and at the same time,the upsurge of national garbage classification has gradually arisen.However,a large number of wastes are put into the garbage can together without being classified,and finally sent to the garbage dump for garbage classification and disposal.However,at present,most of the methods of garbage classification are still backward manual classification.The efficiency of manual classification is very low and there will be classification errors,so we can use other technologies to help the classification of garbage.As a popular technology because of the improvement of hardware devices in recent years,deep learning has been gradually applied to various fields,such as speech recognition,image recognition,face recognition and so on.Therefore,in order to improve the efficiency of garbage classification and reduce the workload of manual classification or completely replace manual classification.So,this paper proposes a study and implementation of garbage classification and recognition methods based on deep learnin,so as to achieve the purpose of garbage classification and recognition with deep learning method.The specific research work of this paper is as follows:(1)Firstly,this paper investigates the current situation of garbage classification and deep learning at home and abroad.SSD,Faster R-CNN and YOLOv3 are selected to analyze and compare because of the two-stage detection algorithm and one-stage detection algorithm which are most used nowadays.According to the advantages and disadvantages of the three algorithms,two mainstream algorithms,Faster R-CNN and YOLOv3,are finally selected for garbage classification and recognition.Making the garbage data set needed in this paper,including the making of training set and test set,data annotation and data expansion because of the lack of data.This data set provides the data base for the later garbage classification and recognition.(2)This paper proposed a method of garbage classification and recognition based on Faster R-CNN.Firstly,build an experimental platform for garbage classification and recognition with Faster R-CNN,and design the experiment.Use the garbage training set made before to train in the Faster R-CNN model for many times under different parameters and test on different test sets.And the training process and test results are analyzed.Finally,the best training weight in this experiment is selected to prepare for the later hardware implementation.(3)This paper also proposed a method of garbage classification and recognition based on YOLOv3.Build an experimental platform for domestic waste classification and recognition with YOLOv3,and carry out the experimental design.Use the domestic waste training set made before to train in the two models,Darknet-53 model and yolov3-tiny model,for many times under different parameters.Then test the weight of the training completed on different domestic waste test sets for many times,and analyze the training process and test results.Finally,the training weights of the two training models which have the best effect on garbage classification and recognition in this experiment are selected to prepare for the later hardware implementation.(4)After the training,the weights of Faster R-CNN and YOLOv3 are implemented on the Jetson Nano embedded platform.And the results of using the Jetson Nano embedded platform are shown and analyzed.After the research in this paper,when the learning rate is 0.00025,the Faster RCNN model trained many times has the highest accuracy for garbage classification and recognition.The accuracy can reach 89.19% and the missed rate is 2.97%.Using the Darknet-53 model in YOLOv3 to classify and recognize the domestic garbage,when the learning rate is 0.001,the Darknet-53 model trained many times has the highest accuracy for garbage classification and recognition.The accuracy rate can reach 91.28% and the missed rate is 2.27%.Using yolov3-tiny model in YOLOv3 to classify and recognize the domestic garbage,when the learning rate is 0.001,yolov3-tiny model trained many times has the highest accuracy for garbage classification and recognition.The accuracy rate is 81.86% and the missed rate is 11.57%.Through comparison,it is found that the Darknet-53 model has the best classification and recognition effect without considering the recognition speed and other conditions.The Faster R-CNN model also has a good effect,but the yolov3-tiny model has the worse effect compared with the above two.However,when running on the Jetson nano embedded platform,the recognition speed of Darknet-53 model and Faster RCNN is very slow.Only yolov3-tiny model can run smoothly and achieve real-time detection.