TSRes-YOLO: An accurate and fast cascaded detector for waste collection and transportation supervision

Abstract

This paper addresses the problem of frequently mixed loading in the waste collection and transportation process, which leads to degradation of the waste classification effect. A supervision method based on a cascaded detector combined with ResNet18 and YOLOv5 for waste collection and transportation is proposed. First, the improved YOLOv5s is used to identify the location and states of the garbage cans. By replacing batch normalization in the backbone with representative batch normalization, the specific features are enhanced with a simple and effective feature calibration scheme and produce a more stable feature distribution. Introducing CBAM into the CSP module of YOLOv5 helps the network pay more attention to special regions. The Ghost convolution effectively reduces the number of parameters and faster the speed without affecting the performance of the model. Moreover, the classifier, improved ResNet-18, classifies the color of garbage cans detected by YOLOv5. SE-Net is added at the beginning of the residual structure, and ResNet-D structure is used to improve the accuracy of the network at the same time. Finally, a database for waste classification and collection supervision is constructed based on the images provided by the relevant city department. Then, a series of experiments are performed on the database. The experimental results show that the improved YOLOv5 network reaches an mAP of 98.0%, the accuracy of the improved ResNet18 reaches 98.4% and the accuracy of the cascaded detector reaches 90.93% with a speed of 43.2 ms/it, which meets the requirements for urban garbage collection and transportation supervision.