AGDN: a lightweight interference-resistant model for photovoltaic panel surface damage detection under complex field conditions

Abstract To address the challenges of "abnormal feature dilution" and "interference amplification" in photovoltaic panel surface damage detection under complex field conditions, this paper proposes the Alignment-Gating-Discrimination Network (AGDN), a lightweight object detection model built upon YOLO11n. First, an enhanced dynamic upsampling module, SA-DySample, is developed by extending the DySample framework with a Lightweight Channel-Aware (LCA) mechanism, an Amplitude-Aware Stabilization (AAS) mechanism, and a Flexible Boundary Mechanism (FBM). While suppressing background noise via channel recalibration, SA-DySample incorporates magnitude-dependent damping constraints and mathematical boundary buffers to effectively curb sampling drift, alleviate edge truncation distortion, and prevent gradient lock-up. Subsequently, the SCSA attention module is deployed after P4 feature concatenation as a pre-gating unit to mitigate the co-activation and amplification of multi-source interference while preventing the abnormal dilution of weak features. Post-fusion, a content-adaptive refinement unit, C3K2_OD, is constructed at the P4 and P5 nodes by integrating omni-dimensional dynamic convolution into the C3K2 structure, transcending static parameter-sharing constraints to precisely differentiate weak damage responses from confounding noise. Furthermore, a bypass-residual-free MobileNetv1_n backbone is adopted to minimize memory access cost and computational overhead. Experimental findings demonstrate that AGDN achieves a Precision, Recall, and F1-Score of 92.93%, 90.39%, and 91.64%, alongside mAP50, mAP75, and mAP50-95 metrics of 94.89%, 85.40%, and 80.82%, respectively. Compared to the baseline model, the computational cost, parameter count, and model size are compressed by 36.36%, 18.33%, and 16.98%, respectively. Moreover, while maintaining a comparable Recall, the Precision, F1, mAP50, mAP75, and mAP50-95 are significantly enhanced by 4.78, 2.17, 1.05, 2.52, and 3.45 percentage points, respectively. Additionally, the model yields a 35.1FPS inference speed and a mere 28.5ms latency when deployed on the NVIDIA Jetson Orin Nano (FP16). Evaluations on two public datasets demonstrate AGDN's competitive cross-dataset adaptability, particularly in Recall, F1-score, and mAP50, supporting its applicability to photovoltaic inspection scenarios.