Unsupervised Semantic Segmentation for Solar Cell Defect Detection Using SAM and Feature Fusion

This study investigates the application of the Segment Anything Model (SAM) proposed by Meta AI for unsupervised semantic segmentation of surface defects in solar cells using the Electroluminescence Photovoltaic (ELPV) dataset. SAM, combined with feature extraction and fusion techniques such as VGG16 feature maps generation and ORB detection, was utilized to generate segmentation masks without any further fine-tuning or training, but simply with points prompt. This research demonstrates that integrating ORB with pretrained VGG16 model extracting deep image features significantly improves the accuracy of segmentation masks generated by SAM, making it a promising approach for further solar defect detection study. Average evaluation confidence score of automatically generated mask increased from 0.59693 to 0.67698.