SolarVQA: A Benchmark for Visual Question Answering in Photovoltaic Defect Inspection

Dataset Summary

SolarVQA is a structured Visual Question Answering dataset built from expert-annotated electroluminescence (EL) images of silicon solar cells. It contains 130,712 QA pairs across 16,339 images spanning eight complementary question types designed to probe defect existence, counting, type identification, severity, localisation, co-occurrence, and spatial distribution.

Dataset Structure

Each QA record contains:

• question_id: unique UUID
• filename: EL image filename (images available from the original source — see below)
• question_type: one of eight types
• question: question string
• answer: ground-truth answer (open-ended string)
• choices: list of multiple-choice options
• correct_choice_idx: integer index into choices
• evidence: JSON dict with the annotation evidence used to derive the answer
• img_width, img_height: image dimensions in pixels
• num_regions: number of annotated defect regions in the image

Splits

Split	Images	QA Pairs	Defect Ratio
Train	11,799	94,392	68.9%
Val	1,320	10,560	69.2%
Test	3,220	25,760	69.5%

Question Types

Type	#Choices	Random Baseline
Existence	2	0.500
Counting	4	0.250
Type Identification	11	0.091
Severity	5	0.200
Localisation	10	0.100
Co-occurrence	2	0.500
Comparative (H)	4	0.250
Comparative (V)	4	0.250

Defect Classes (9 total)

Contact: Contact_Corrosion, Contact_FrontGridInterruption, Contact_NearSolderPad
Crack: Crack_Closed, Crack_Isolated, Crack_Resistive
Interconnect: Interconnect_BrightSpot, Interconnect_Disconnected, Interconnect_HighlyResistive

Source Images & Data Preparation

EL images and their annotations are NOT included in this repository. They must be downloaded from the UCF-EL-Defect dataset.

Dataset Cleaning & Splitting Pipeline

To prepare the raw UCF-EL-Defect dataset for SolarVQA generation, this repository includes a preprocess.py script. This script ensures the integrity and balance of the base data before VQA pairs are generated by performing the following operations:

• Data Filtering: Removes files missing from the CSV, enforces image naming conventions, resolves shape/attribute conflicts, and cleans exact duplicate CSV rows.
• Class Pruning: Completely removes noisy or unwanted classes (Unknown, Contact_BeltMarks).
• Targeted Redistribution: Strategically redistributes the Interconnect_Disconnected class across the Train, Val, and Test sets to ensure a balanced distribution for rare defects.
• Reproducible Splits: Derives a stable Validation split (10% of the original Train set) and proportionally assigns new, unallocated images across the final base folders.
• Clean Outputs: Physically routes valid images into train, val, and test directories and generates freshly re-indexed CSV annotation files for each split.

Cross-Split Consistency

The files severity_bins.json and defect_class_vocab.json (in data/train/) are derived exclusively from the training split and must be used unchanged for validation and test to guarantee consistent answer vocabularies.

License

CC BY 4.0. QA annotations are released under this license. Original EL images retain their original license from the source dataset.