ACFNet: Asymmetric Cross-Layer Fusion Network with Semantic Context Enhancement for Photovoltaic Defect Detection

Photovoltaic (PV) cell defects, such as cracks and star cracks, often arise during manufacturing and operation and may adversely affect the electrical performance and long-term reliability of photovoltaic cells. Detecting these defects in photovoltaic defect images is challenging due to complex background textures, scale variations, and structural discontinuity of defects. To address these challenges, we propose an asymmetric cross-layer fusion network (ACFNet) for photovoltaic cell defect detection. Specifically, we design a spatial-channel modulating block (SCMB) into the backbone to suppress background interference through joint spatial and channel modulation. We propose an asymmetric cross-layer fusion (ACF) strategy to handle scale variations by performing multiscale feature aggregation via asymmetric cross-layer connections. Furthermore, we develop a semantic context enhancer (SCE) in the detector heads to enhance the representation of discontinuous defect structures by combining large-kernel convolutions with parallel attention mechanisms. Experimental results on the PVEL-AD and PV-Multi-Defect datasets show that ACFNet achieves mAP@0.5 values of 0.933 and 0.884 and mAP@0.5:0.95 values of 0.652 and 0.598, respectively, with 3.0 M parameters and a measured throughput of 75.7 FPS on an NVIDIA A100 GPU, equivalent to an average processing time of approximately 13.2 ms per image.