MS-YOLOv11: A multi-scale feature fusion network for real-time rooftop photovoltaic detection from UAV images

Abstract

With the rapid proliferation of distributed solar energy, Unmanned Aerial Vehicles (UAVs) have emerged as indispensable tools for rooftop photovoltaic (PV) inspection. However, practical deployment is hindered by hierarchical challenges, including extreme scale variations of PV targets, complex background interference, and the limited computational resources of edge devices. To address these issues, this study proposes a lightweight yet high-performance detection network based on the YOLOv11 architecture. First, to mitigate the missed detection of small-scale targets, a high-resolution P2 detection layer is constructed to preserve fine-grained features from shallow layers. Second, a Large Selective Kernel Attention (LSKBlock) mechanism is integrated into the backbone, enabling dynamic receptive field adjustment to enhance feature extraction capabilities in cluttered environments. Finally, the G3Ghost module is introduced to counterbalance the computational overhead of the multi-scale structure. This module leverages linear transformations to reconstruct redundant features, significantly compressing the model volume without compromising precision. Experimental results on a specialized UAV-based rooftop PV dataset demonstrate superior performance. Compared with the baseline model, the proposed method achieves a precision of 96.32%, a recall of 95.74%, and an mAP@0.5 of 98.5%. Notably, the model is exceptionally lightweight, with only 2.04 M parameters and a 4.5 MB storage footprint, while maintaining a high inference speed of 115 FPS. These metrics indicate that the proposed approach achieves an optimal trade-off between detection accuracy and operational efficiency, providing robust technical support for mobile deployment on UAV platforms.