Transplantation of retinal pigment epithelial (RPE) sheets derived from human induced pluripotent cells (hiPSC) is a promising cell therapy for RPE degeneration, such as in age-related macular degeneration. Current RPE replacement therapies, however, face major challenges. They require a tedious manual process of selecting differentiated RPE from hiPSC-derived cells, and despite wide variation in quality of RPE sheets, there exists no efficient process for distinguishing functional RPE sheets from those unsuitable for transplantation. To overcome these issues, we developed methods for the generation of RPE sheets from hiPSC, and image-based evaluation. We found that stepwise treatment with six signaling pathway inhibitors along with nicotinamide increased RPE differentiation efficiency (RPE6iN), enabling the RPE sheet generation at high purity without manual selection. Machine learning models were developed based on cellular morphological features of F-actin-labeled RPE images for predicting transepithelial electrical resistance values, an indicator of RPE sheet function. Our model was effective at identifying low-quality RPE sheets for elimination, even when using label-free images. The RPE6iN-based RPE sheet generation combined with the non-destructive image-based prediction offers a comprehensive new solution for the large-scale production of pure RPE sheets with lot-to-lot variations and should facilitate the further development of RPE replacement therapies.