Photovoltaic (PV) panel defect classification is essential for reliable solar energy systems, particularly in industrial inspection. This study proposes a lightweight ConvNeXt-based architecture with a hybrid feature aggregation mechanism that combines Global Average Pooling (GAP) and Global Max Pooling (GMP), followed by a compact classification head for multiclass defect classification. Unlike standard pooling strategies, the proposed aggregation enhances feature representation by capturing both global distribution and salient responses. The proposed model is evaluated through ablation studies and compared with baseline configurations under a unified training pipeline. Experimental results show that the proposed architecture achieves a weighted F1-score of approximately 0.96, outperforming baseline ConvNeXt variants and demonstrating improved feature representation capability. These findings indicate that effective yet simple feature aggregation strategies can significantly enhance classification performance while retaining reasonable inference speed for industrial applications, making the model suitable for real-world PV inspection applications.
Berkay Eren (Fri,) studied this question.