Fault-Level Grading of Photovoltaic Cells Employing Lightweight Deep Learning Models

Table 11

Comparative analysis for electroluminescence image database classification.

Comparison parameters	Study
	[38]		[40]	Proposed method
Database (original)	2,624		2,624	2,624
Data division (%)	75–25 (train-test)		80–20 (train-test)	70-15-15 (train-val-test)
Training samples after augmentation	196,800		Not mentioned	6,000
Features	SIFT, SURT, KAZE, HOG, PHOW		×	GLCM, LBP	×
Classifier	SVM	CNN	VGG-based CNN	Deep ANN	Customized CNN
Binary classification accuracy	82.44%	×	93.02%	92.1%	94.3%
Binary classification accuracy with 0.5 as threshold	×	88.42%	×	84.8%	89.3%
Multiclassification accuracy	×	×	×	76.1%	83.5%