MCSFI: A detection network for photovoltaic panel defect detection of multi-scale content-aware feature integration

Abstract

The development of photovoltaic (PV) panel systems not only mitigates pollution caused by fossil fuel combustion but also addresses the growing global demand for sustainable energy. Defect detection in PV panels is critical to ensuring the reliable operation of PV power systems. However, existing methods for defect detection face challenges in balancing computational resource efficiency with detection accuracy. To address these limitations, this article proposes the Multi-Scale Content-Aware Feature Integration (MCSFI) network model, which achieves enhanced detection performance while maintaining a lightweight design. First, the article introduces the SARepVGG module, integrated into both the Backbone and Neck networks, to strengthen the model's ability to represent defect-related features. Second, the article designs a Multi-Scale Context-Aware Feature Enhancement (MFCARAFE) module, which processes outputs from multiple convolutional layers in order to comprehensively aggregate defect features across different scales. This significantly improves detection accuracy for PV panel defects of varying sizes. Third, the article proposes the Adaptive Input Feature Integration Convolution (AIFIC) module, which combines adaptive input feature calibration with dual-path convolutional technique to enhance the model's adaptability to complex scenarios and generalization capabilities. Extensive experiments on the PVEL-AD dataset and Dataset A validate the effectiveness of our approach. Compared with the baseline model, the proposed MCSFI model achieves a 1.3% improvement in mAP on the PVEL-AD dataset while reducing the model weight size by 6.1%. Similar performance achievements are observed on Dataset A. These results demonstrate that our method successfully balances multi-scale defect detection accuracy with computational efficiency, offering a novel solution for practical PV panel defect inspection.