Garbage Detection And Classification Based On Deep Learning

With the influx of urban population and the continuous improvement of living standards,the types and quantities of garbage generated are increasing,which also brings serious garbage disposal problems.Although there are currently methods that use deep learning object detection for garbage detection and classification,they don’t achieve the effect of accurate and fast identification of garbage,and it has not yet met the requirements of practical application in the industrial field.Therefore,the research purpose of this paper is to improve the accuracy and speed of model recognition at the same time,which is convenient for deployment to embedded devices.The main research contents and innovations are as follows:(1)Aiming at the problems of YOLOv5 algorithm network to insufficient ability of extract small garbage target features,insufficient utilization of feature fusion information,and low detection accuracy,in this paper,the garbage detection and classification algorithm Focal＿ECA＿SPP-Fast-YOLOv5(FES-YOLOv5)based on attention mechanism and multifeature fusion is proposed.Firstly,multiple garbage datasets are integrated and data augmentation and data resampling are used to solve the problem of sample imbalance between garbage categories,so as to improve the generalization and robustness of the model?Secondly,the anchor clustering algorithm is improved by fusing genetic algorithm to obtain the global optimal anchors size suitable for the current garbage dataset? Then the lightweight channel attention network is added to the adjacent two layers of the feature extraction network downsampling to enhance the feature extraction ability of the network and improve the detection effect of small garbage targets? Finally,the single-pooled kernel serial network is introduced to fully integrate the feature information to further improve the feature fusion ability of the model.Compared with the YOLOv5 and YOLOv5-4DWGA garbage detection algorithm,the m AP increases by 0.0488 and 0.042,respectively.The detection results of small garbage targets and multi-garbage targets in actual scenarios show that the algorithm can reduce the false detection rate and false detection rate.(2)Aiming at the problems that FES-YOLOv5 model is large in size,has redundant parameters and is slow to detect when deployed on embedded devices in this research scenario,in this paper,the model compression algorithm for joint pruning is proposed.Firstly,the global threshold pruning method of BN layer is used to implement channel pruning for all network layers except the Bottleneck structure,the best regular coefficientλ=0.001 suitable for this model is found by sparse training? Secondly,add a convolution operation to the shortcut in the Bottleneck structure,use the convolution kernel weight pruning method to prune the main branch and shortcut and ensure that the output dimensions are consistent? Finally,set different channel pruning ratios and the number of channels to achieve model compression.In the case of channel pruning by 20%,the model size is reduced from 13.91 M to 11.10 M,the detection speed is increased from 9.2FPS to11.5FPS,and the m AP is reduced by 0.0382.(3)Based on the design principles of convenient use,reasonable layout and clear interface,this paper implements a garbage detection and classification system on embedded devices based on Vue front-end and Flask back-end frameworks.The system includes registration,login and garbage image detection interfaces,function implementation of front end,back end and detection interface AIDetector,and the test results are presented through the visual interface.This system has simple interface and perfect functions.