Zhaoyang Wang

Photovoltaic wafer manufacturing demands near-zero defects, yet anomaly detection models often fail to generalize across manufacturing lines due to domain shift caused by varying imaging devices and process conditions. This letter presents Projection Comparison and Reconstruction (PCR), a framework for cross-domain anomaly detection. Patch-based Domain Fusion mixes source and target images via grid-wise stochastic patch replacement, explicitly increasing cross-domain distribution overlap. Dual-Teacher Knowledge Distillation combines a frozen domain-agnostic teacher with a trainable adaptive teacher to obtain domain-invariant features for robust reconstruction. PCR achieves state-of-the-art performance, including an 8.07-point image-level AUROC gain on ELAD.

The vision foundation model, relying on large-scale pre-training, has advanced image comprehension capabilities and excels in general scenarios. However, its performance remains suboptimal in specialized tasks, such as photovoltaic (PV) cells defect detection. This limitation stems from the models’ lack of domain-specific prior knowledge. To address this, we propose a two-stage pre-training framework comprising fine-grained feature autoencoding (FFA) and pseudo-box contrastive learning (PCL), which leverages extensive unlabeled raw images to inject domain expertise into the model. First, we investigate the fine-grained feature autoencoder, which trains a detail-sensitive vision transformer (ViT) backbone by reconstructing the histogram of oriented gradients (HOG) of masked images. Second, we pre-train the detection head through contrastive learning. Using selective search (SS) to generate pseudo-boxes, we treat paired boxes from two augmented views of an image as positive samples. The abundant unsupervised pseudo-boxes optimize the detection head’s local representation and localization capabilities. Finally, we fully fine-tune the model with labeled images. Based on this methodology, we build the cross-scenario photovoltaic defect detector (CPDD). The experimental results demonstrate that CPDD achieves state-of-the-art (SOTA) mAP50 scores on three benchmarks, outperforming detectors pre-trained on the COCO dataset as well as those specifically designed for PV defect detection.

Due to the advantages of safety and economy, it has become a trend to use unmanned aerial vehicles (UAVs) instead of humans to inspect high-voltage transmission lines. Considering the manual inspection process and the few-shot learning, a two-stage method for insulator anomaly detection is proposed. In the first stage, a positioning-restoration-cropping method is discussed for insulator string detection and processing. In the second stage, an insulator anomaly detection model called a multi-scale feature reweighting (MFR) network is built. With the help of few-shot object detection, the detection of five kinds of anomaly insulator caps, such as falling off, breakage and ablation is realized. The mean average precision (mAP) of the proposed method is 88.76%.