Anchor-Free vs. Anchor-Based Detectors for Multi-Scale Thermal Anomaly Detection in Solar PV Fields: An Empirical Study

This paper’s key contribution is a systematic empirical comparison of anchor-based vs. anchor-free object detectors for multi-scale thermal hotspot detection in solar PV fields, using a real UAV-acquired dataset (3,116 thermal images, Kelantan and Pulau Pinang, Malaysia) under a unified training protocol. It demonstrates that anchor-free models (YOLOv8/v10/v11) consistently and substantially outperform anchor-based detectors (Faster R-CNN, SSD, RetinaNet) — best anchor-free mAP0.5 of 95.78% and mAPsmall of 50.24% vs. best anchor-based 87.63% and 35.62%, respectively — directly attributing this to anchor-free dense regression avoiding the anchor-to-object mismatch caused by PV arrays’ extreme aspect ratios and tiny hotspot sizes. It also quantifies the accuracy-efficiency trade-off for edge/UAV deployment, providing an empirical basis for choosing detection architectures in real-time PV inspection systems — directly relevant groundwork for your own YOLOv9-based hotspot detection work.