Research On Garbage Classification And Detection Methods In Complex Scenes

Domestic garbage is a misplaced resource.An effective garbage classification strategy can not only protect the environment,but also realize the recycling and reuse of garbage,which has high economic benefits.The classification and treatment of garbage in my country’s garbage treatment plants is still in the manual sorting stage,which not only causes problems such as large labor and low sorting efficiency,but also misclassification of garbage.In recent years,with the continuous development of deep learning technology,the accuracy of target detection and recognition has been effectively improved,and deep learning has the ability to learn independently.Therefore,this paper proposes a garbage classification and detection model based on improved YOLOv3(You Only Look Once).The research contents of this paper are as follows:(1)Aiming at the problem of relatively single garbage classification and detection data set and insufficient algorithm robustness in complex scenarios,this paper adopts scale normalization and data enhancement strategies for garbage classification detection data set to expand the data set(2)Aiming at the problem of large classification and positioning errors of the Darknet53 network model,this paper studies the addition of a set of residual modules after the last set of residual blocks of the network to perform feature fusion,so that the improved network can generate 4 types Feature maps of different scales are used to detect targets of different scales in the image.In addition,due to the unstable results of the K-means clustering algorithm of YOLOv3’s prediction box generation method,this paper uses a more advanced K-means++ clustering algorithm to generate the prediction box;(3)For the garbage classification and detection data set,the garbage size is small and the distribution is dense.In this paper,the distance IoU(Distance IoU,DIoU)loss is used as the loss function of the prediction box regression,and the distance and overlap between the real box and the prediction box are comprehensively considered.In the prediction frame screening stage,DIoU-NMS technology is used to reduce the confidence of the prediction frames of two different objects,reduce the occurrence of missed detection,and improve the detection accuracy of the model.The experimental results show that the improved algorithm in this paper achieves93.14% m AP and the detection speed is 45.76 fps,which verifies the effectiveness of the detection model.