| In recent years,as the living standards of people in our country rise year by year,the amount of trash generated has also exploded,making the disposal of trash a worldwide problem.The Chinese government’s concern about environmental issues has gradually deepened,and various trash sorting policies and laws and regulations have been continuously introduced to supervise citizens’ trash sorting.Under the current framework of trash sorting laws,regulations and policies,this article uses artificial intelligence technology to study the automatic detection of trash targets.With the powerful computing power of computers,it is possible to automatically identify trash categories and trash location information,thereby promoting Implementation of the trash sorting policy.This subject mainly studies the acquisition of trash images in the intelligent trash inspection vehicle and the realization of trash detection algorithms.The trash detection algorithm also includes trash sorting and target positioning algorithms,which can perform multiple trash detection in an image.Judging the category,obtaining the location information of each trash target at the same time,and automatically frame selection on it.The main contents of this research work include the following aspects:The main research of this article is as follows:(1)Based on Huawei’s trash sorting datasets,I labeled the datasets using in trash detection algorithms for training and testing by myself;(2)The classical image feature extraction algorithm and the convolutional neural network feature extraction algorithm have been studied theoretically and experimentally.The proposed convolutional neural network recognition accuracy rate is 11.82% higher than the SIFT algorithm with the highest accuracy rate of trash sorting and recognition in the classic feature extraction algorithm;(3)The backbone network of the trash detection algorithm is studied,and the DL-TS deep learning model is innovatively proposed on the basis of Dense Net169.By introducing two different attention mechanisms,the recognition accuracy is increased from 91.62% to 94.35%.By adding the variability convolution module,the recognition accuracy was further improved,and finally reached 97.63%;(4)The trash detection algorithm was researched,and the core structure of DL-TS was innovatively added on the basis of YOLOv4,and the DLTS-YOLOv4 trash detection algorithm was proposed,which increased the m AP,m IOU and FPS values from 80.37%,88.75% to 82.67% and 91.36%.Then,on this basis,the Kmeans method is used to modify the size of the anchor frame,so that the m AP and m Io U values of the algorithm are improved again on the original basis,reaching 83.47% and 92.75%.Finally,in response to the problem of missed detection of small targets,a small target recognition anchor frame was added,which once again improved the recognition effect.The final m AP value reached 84.91%,while the m Io U value was 93.64%,and the FPS value has been maintained at 34 or 36 frames per second. |
