Related Experiment Video
Updated: Jun 27, 2026
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
Evaluating Convolutional and Transformer Architectures for Photovoltaic Defect Classification via Electroluminescence
Seda Bayat Toksöz1, Gültekin Işık1, Gökhan Şahin2,3
1Department of Computer Engineering, Iğdır University, Iğdır 76000, Türkiye.
Sensors (Basel, Switzerland)
|June 26, 2026
Summary
This study benchmarks deep learning models for photovoltaic defect inspection using electroluminescence imaging. ConvNeXt-T shows superior performance in identifying defects across various cell types and tasks.
Area of Science:
- Materials Science
- Electrical Engineering
- Computer Science
Background:
- Electroluminescence (EL) imaging is crucial for photovoltaic (PV) defect detection.
- Comparing deep learning (DL) models for PV defect inspection is challenging due to inconsistent datasets and methodologies.
Purpose of the Study:
- To establish a controlled, image-level benchmark for evaluating six DL architectures on PV defect inspection.
- To provide a standardized comparison across diverse tasks and cell types.
Main Methods:
- A proprietary dataset of 20,000 single-cell EL images was used.
- Six architectures (ConvNeXt-T, ViT-B/16, DeiT-B/16, Swin-T, DenseNet121, MobileNetV3-Large) were evaluated.
- Identical preprocessing, augmentation, training, and five-fold cross-validation were applied across 150 runs.
Main Results:
- ConvNeXt-T achieved the highest mean macro-F1 score (93.12%) with fewer parameters than ViT/DeiT models.
- On a four-class polycrystalline task, ConvNeXt-T reached 84.94% macro-F1, outperforming DenseNet121 (70.08%) and MobileNetV3-Large (59.43%).
- Lightweight CNNs exhibited conservative missed-defect behavior, particularly for surface degradation and cracks.
Conclusions:
- ConvNeXt-T demonstrates state-of-the-art performance for PV defect inspection via EL imaging.
- The controlled benchmarking provides reliable cross-validation evidence for model selection.
- Future research should focus on module-level grouped validation for enhanced defect detection.