Hotspot defect detection for photovoltaic modules under complex backgrounds

Hotspot defect detection (HDD) of photovoltaic (PV) modules is one of the daily inspections of PV power stations. It aims to detect hotspot defects from the infrared images(IFIs), which are captured by the unmanned aerial vehicles at about 20 ms. The backgrounds in the IFIs are complex, which results in the difficulties of detecting hotspots in PV modules, especially the tiny-size hotspots. Therefore, a segmentation-before-detection method is proposed for HDD in this paper. In specific, the regions of PV modules in the IFIs are first extracted by an improved semantic segmentation model, and then hotspot defects are detected from the segmented regions by a developed object detection model. The semantic segmentation model is named Attention DeepLab, which has been developed by an attention module. And, the object detection model is derived from YOLOv5s. Three optimization schemes are proposed to increase the detection accuracy for tiny-size hotspot defects. The schemes are: (1) appending a prediction head for tiny hotspot defects in the prediction network, (2) revising the path aggregation network in the feature fusion network by merging multiple-scale feature maps to enhance the semantic information, and (3) applying efficient channel attention module to eliminate aliasing in the feature fusion network. Experimental results demonstrate the effectiveness of the proposed method. The mean intersection over union of the semantic segmentation model of PV modules is 97.8%, and the average precision of the HDD is 89.6%.