DRSE-YOLO: Efficient and Lightweight Architecture for Accurate Waste Detection

Abstract

This paper introduces DRSE-YOLO, an efficient waste detection model designed to address detection accuracy and lightweight design challenges. The RCCA module in the model's neck enhances multi-scale feature representation, thereby improving detection performance. The DySample module optimizes upsampling through adaptive point-sampling, reducing computational demands and improving resource efficiency. The Slim-Neck module is applied to select convolutional layers and C2f modules to streamline the model and enhance computational efficiency. The ECC-Head integrates asymmetric depth convolution, point convolution, and an attention mechanism, balancing accuracy with reduced parameters and computational load. Evaluated on a custom dataset comprising 46 waste classes and approximately 25,000 images, DRSE-YOLO achieves significant improvements over YOLOv8n, including a higher [email protected] (+1.59%) and [email protected]:95 (+2.08%), alongside a reduced parameter count (2.43 M vs. 3.2 M) and GFLOPs (5.8 vs. 8.2, a 24.4% reduction). These results underscore DRSE-YOLO's efficiency and accuracy.