An improved YOLO11 algorithm for scenic garbage detection based on machine vision

Abstract

Driven by the rapid expansion of smart tourism, real-time garbage detection leveraging existing surveillance infrastructure has become a critical task for environmental management. However, garbage targets in scenic areas are often characterized by extremely small sizes and are embedded in complex natural backgrounds, leading to high missdetection rates. This paper proposes Scenic-YOLO11, an improved YOLO11 algorithm. We introduce the RCS-C3k2 module to mitigate feature loss during downsampling through structural re-parameterization. An Efficient Multi-Scale Attention (EMA) mechanism is integrated into the Neck network to suppress background noise. Furthermore, we design a Dual-Branch Head that couples high-resolution P2 features with deep semantic P5 features to enhance the detection sensitivity of tiny objects. To accelerate convergence and refine localization, an Inner-MPDIoU loss function is adopted. Experimental results on the custom SG-Video dataset demonstrate that Scenic-YOLO11 achieves an mAP50 of 87.2% (an 8.0% improvement over YOLO11n) and an mAPs of 35.8%, while maintaining a high inference speed of 98 FPS with only 3.5M parameters. This empirical study proves the model's efficacy for real-time edge deployment in complex scenic environments.