Integrating Superpixel Segmentation with YOLOv8 for Robust Detection of Densely Packed Electricity Meters

Detecting densely arranged electricity meters is challenging due to severe overlap, blurred boundaries, and limited spacing, which hinder conventional detectors. To address this, we propose SPG-YOLO, a Superpixel-Guided YOLOv8 framework that improves spatial awareness and robustness. It adopts a dual-stage superpixel modeling strategy: the first stage extracts global structure via convolutional encoders and a lightweight context module using downsampled macro-patches, while the second stage captures fine-grained local features from micro-patch sequences and image crops. These multi-scale features are fused to generate boundary-aware superpixel maps, which are early-fused with RGB images and input into a modified YOLOv8. This explicit structural guidance enhances detection in dense, cluttered scenes. Experiments on a real-world meter dataset show that SPG-YOLO outperforms standard YOLOv8 and other baselines in both accuracy and robustness.