A Large-scale Evaluation of Pretraining Paradigms for the Detection of Defects in Electroluminescence Solar Cell Images

TL;DR

This work tackles defect detection in electroluminescence (EL) solar-cell images under limited labeled data by benchmarking a broad spectrum of pretraining paradigms: supervised (COCO, ImageNet), self-supervised (SimCLR, MoCov2 on ImageNet and EL), and semi-supervised (CCT, U2PL). It systematically analyzes both out-of-distribution and in-distribution pretraining, revealing that supervised COCO, self-supervised ImageNet, and semi-supervised CCT achieve statistically equivalent $mIoU$ on SCDD, with U2PL underperforming. The study introduces a large unlabelled EL image dataset (22,000 publicly released) and a 642-image, ground-truth semantic segmentation benchmark, achieving a new state-of-the-art and highlighting better handling of underrepresented defects in some regimes. The findings suggest that domain-tailored self-supervision remains challenging for EL data, while large-scale OOD pretraining provides robust, deployable gains, and it provides datasets to spur further research in this domain.$mIoU$ and $wIoU$ are used to quantify segmentation performance. The work thus informs practical model selection for SCDD and motivates future domain-specific self-/semi-supervised technique development.

Abstract

Pretraining has been shown to improve performance in many domains, including semantic segmentation, especially in domains with limited labelled data. In this work, we perform a large-scale evaluation and benchmarking of various pretraining methods for Solar Cell Defect Detection (SCDD) in electroluminescence images, a field with limited labelled datasets. We cover supervised training with semantic segmentation, semi-supervised learning, and two self-supervised techniques. We also experiment with both in-distribution and out-of-distribution (OOD) pretraining and observe how this affects downstream performance. The results suggest that supervised training on a large OOD dataset (COCO), self-supervised pretraining on a large OOD dataset (ImageNet), and semi-supervised pretraining (CCT) all yield statistically equivalent performance for mean Intersection over Union (mIoU). We achieve a new state-of-the-art for SCDD and demonstrate that certain pretraining schemes result in superior performance on underrepresented classes. Additionally, we provide a large-scale unlabelled EL image dataset of $22000$ images, and a $642$-image labelled semantic segmentation EL dataset, for further research in developing self- and semi-supervised training techniques in this domain.

Figures (7)

• Figure 1: EL images of multi-crystalline (top) and mono-crystalline (bottom) silicon solar cells
• Figure 2: Number of images containing each defect class (top) and feature class (bottom)
• Figure 3: EL and ground truth from UCF fioresi2022automated (top) and the CSIR dataset pratt2023 (bottom)
• Figure 4: mIoU by class type (extrinsic defect and intrinsic feature).
• Figure 5: wIoU by class type (extrinsic defect and intrinsic feature).