PPS-YOLO: improved YOLOv11 algorithm for small target defects detection of photovoltaic cell modules

Abstract Reliable detection of defects in photovoltaic (PV) cell modules is essential for maintaining high efficiency and ensuring long-term durability. To overcome several difficulties, such as complex background noise, diverse defect categories, and the presence of small-scale flaws, this study introduces PPS-YOLO, an enhanced defect detection framework built upon YOLOv11. To strengthen the performance on tiny targets, a lightweight multiscale pinwheel convolution (LMSPConv) module is developed and applied to the backbone’s P1 and P2 layers. Additionally, the C3k2_PKI module is incorporated into the backbone, whose multiscale convolutional design and contextual attention mechanism improves the network’s sensitivity to defects of different scales in electroluminescence (EL)imagery. To further reduce the computational complexity and parameter count of the model, the SlimNeck structure is adopted for the neck. SlimNeck uses the VoVGSCSP module based on the one-time aggregation strategy to replace the traditional bottleneck structure. This design effectively improves the feature reuse rate and accelerates the inference process. Experiments on the PV EL Anomaly Detection (PVEL-AD) dataset confirm the effectiveness of the proposed method. Compared with the baseline YOLOv11n, PPS-YOLO delivers superior accuracy, achieving a precision of 92%, recall of 90.3%, mAP50 of 95.1%, and mAP50–95 of 68%, representing gains of 0.9%, 3.1%, 1.4%, and 1.3%, respectively.