Solar Cell Defect Detection using Deep-Learning Segmentation with Two-Fold Training

Recently, the applications of Deep Learning (DL) methodologies have been extensively utilized across numerous areas to extract critical solutions from selected databases. The DL Segmentation Tool (DLST) is frequently utilized to extract and assess selected segments from digital images captured using various modalities. This study intends to propose a DLST to analyze the selected database for the identification of defects in solar cells. The proposed DLST encompasses several phases: (i) image acquisition and resizing, (ii) application of an image preprocessing technique to improve defect visibility, (iii) execution of the segmentation process utilizing a selected DL-model to extract the defect, and (iv) comparison of the segmented region with a binary mask and calculation of performance metrics. The created DLST incorporates the VGG-UNet, with VGG19 serving as the model's backbone. This work implemented a two-fold training approach for improved outcomes. The experimental result is validated using selected image metrics, demonstrating that the proposed DLST yields superior values of Jaccard (91.31%), Dice (95.38%), and Accuracy (96.18%) in comparison to other chosen DL models.