Mixture of Experts Detection System for Enhanced Automatic Annotation in Photovoltaic Modules

This paper presents a Mixture of Experts (MoE) detection framework for enhanced automatic annotation in photovoltaic (PV) module inspection. The approach addresses the critical bottleneck of creating high-quality labelled datasets in specialised domains by leveraging the complementary strengths of multiple expert detection models. Rather than relying on a single detection model with inherent limitations, this system combines six expert models (YOLO and RT-DETR across visual, thermal, and multi-spectral data) using a trainable router network that learns optimal weighting strategies for different defect types. The multi-spectral capability is particularly valuable for PV module inspection, where thermal imaging excels at identifying electrical anomalies while visual imaging better captures surface defects. The framework was evaluated using a dataset of 4,111 paired thermal and visual images captured via UAV over a floating PV installation. Experimental results demonstrate that the MoE framework provides more reliable annotations, maintaining competitive detection metrics (mAP:0.645-0.760, IoU:0.810-0.851, GIoU:0.810-0.850) across different imaging modalities, and showing more consistent performance across varying defect types according to individual expert models. The framework accelerates new dataset creation processes and offers a promising approach for developing more effective machine learning applications for PV module fault detection and maintenance.