Research On Garbage Sorting Technology Based On Deep Learning

China produces about 200 million tons of municipal waste every year,most of which can be recycled.At present,the recycling method of garbage in China is mainly through manual sorting.However,due to the huge amount of daily garbage,manual sorting is not only inefficient,but also can damage human health in such a working environment for a long time.Therefore,automatic garbage sorting technology has become a development trend.With the development of computer vision,automatic recognition can be carried out according to the characteristics of the target.However,with the increase of garbage types,the shape of garbage is more and more complex and the placement of garbage is disorderly,the classification effect of traditional machine vision algorithm can not be guaranteed,and the automation system is difficult to be popularized in sorting tasks.As a result,studying the process of exact identification and localization is critical.This paper proposes a garbage sorting system based on deep learning target recognition algorithm to solve the problem of efficient garbage sorting.The main work contents are as follows:1.In-depth study of camera imaging principle and calibration method,use depth camera D435 i and Aubo＿I5 robot to build eye-on-hand vision system according to the experimental environment,and carry out calibration experiments of camera internal and external parameters and hand-eye system combined with calibration tools,so as to make basic preparation for the subsequent accurate grasping.2.Garbage data sets are made by means of collection and random color transformation,and a target detection model based on Yolo V4 is built for garbage classification and location.At the same time,in the post-processing stage,the deflection Angle of garbage is estimated by using ellipse fitting method,experimental results show that the position error ≤3.4mm and the Angle error ≤1.2°,which further improves the grasping accuracy of the system.3.In order to lightweight Yolo V4 and improve detection speed,the main feature extraction network of Yolo V4 was replaced with Efficientnet-B0,and the ordinary convolution in PANet was replaced with the deep deprivable convolution with fewer parameters.When compared to the original network,the number of calculation in the modified model is reduced by 90%,and the detection time of a single image is lowered by 7milliseconds.4.In order to improve the accuracy of small target object detection,a feature fusion method combined with ECA is constructed in PANet to realize weight analysis of feature maps of different channels through cross-channel interaction and improve the feature extraction ability of the model.The mean precision m AP of the improved model increases by 5.77%,and the recall rate and accuracy rate of cigarette butts and batteries are improved.5.Complete the overall design of garbage sorting system,and build Pyqt5 based system interaction interface.Through 300 sorting experiments in the simulated garbage sorting environment,the sorting accuracy of garbage sorting system based on Yolo V4 can reach 94%,which verifies the reliability and practicality of the system.