Efficient Differentiation of Biodegradable and Non-Biodegradable Municipal Waste Using a Novel MobileYOLO Algorithm

Abstract

In the realm of waste management, the accurate identification of biodegradable and non-biodegradable items remains a critical challenge. An advanced real-time object detection method, termed “MobileYOLO”, was proposed, leveraging the strengths of the YOLO v4 framework. The MobileNetv2 network was integrated, and a section of the conventional computation was substituted with depth-wise separable convolutions utilizing the PAnet and head network. To enhance feature expressiveness capabilities during feature fusion, a refined lightweight channel attention mechanism, known as Efficient Channel Attention (ECA), was introduced. The Improved Single Stage Headless (ISSH) context module was incorporated into the micro-object identification branch to broaden the receptive field. Evaluations conducted on the KITTI dataset indicated an impressive accuracy of 95.7%. Remarkably, when compared to the standard YOLOv4, the MobileYOLO model exhibited a reduction in model parameters by 53.12M, a decrease in connectivity size by one-fifth, and an augmentation in detection speed by 85%.