Multi-class photovoltaic panel defect detection using enhanced FCOS

Current defect detection technologies for photovoltaic (PV) panels still face dual challenges: Traditional manual inspection suffers from low efficiency and high costs, failing to meet real-time detection demands in large-scale power plants. While the fully convolutional one-stage (FCOS) object detection algorithm based on deep learning outperforms traditional methods in accuracy and speed, its rectangular bounding box regression mechanism and fixed anchor design struggle to achieve precise contour segmentation and pixel-level localization for irregular defects such as crack propagation and foreign material adhesion. Additionally, the feature pyramid network (FPN) exhibits insufficient multi-scale feature fusion capability, limiting its ability to fully leverage hierarchical feature information. To address these limitations, this paper proposes an improved FCOS-based framework for multi-category defect detection and recognition in PV panels. First, a dual-path feature aggregation block (D-PFAB) is designed to efficiently fuse multi-scale image features by constructing two reverse-direction feature propagation paths, effectively reducing semantic discrepancies between heterogeneous features. Second, a resolution alignment module (RAM) complementary to D-PFAB is introduced, which learns 2D transformation offsets to achieve precise alignment of feature maps across resolutions, thereby enhancing feature fusion quality. Third, an adaptive-weighted multi-task head (AW-MTH) is developed to dynamically adjust task-specific weights for critical features through shared representations, improving both detection accuracy and segmentation performance. Experimental results demonstrate that the proposed method achieves a mean intersection-over-union (mIoU) of 90.8%, with improvements of 2.3% in mean average precision (mAP) and 1.9% in F1-score compared to baseline models. These advancements significantly enhance the detection and segmentation accuracy for complex defects in photovoltaic panels.