Faster and more accurate variation characterizations of semiconductor devices/circuits are in great demand as process technologies scale down to Fin-FET era. Traditional methods with intensive data testing are extremely costly. In this paper, we propose a novel learning-based high-accuracy data prediction framework inspired by learning methods from computer vision to efficiently characterize variabilities of device/circuit behaviors induced by manufacturing process variations. The key idea is to adaptively learn the underlying data pattern among data with variations from a small set of already obtained data and utilize it to accurately predict the unmeasured data with minimum physical measurement cost. To realize this idea, novel regression modeling techniques based on Gaussian process regression and partial least squares regression with feature extraction and matching are developed. We applied our approach to real-time variation characterization for transistors with multiple geometries from a foundry 28nm CMOS process. The results show that the framework achieves about 14x time speed-up with on average 0.1% error for variation data prediction and under 0.3% error for statistical extraction compared to traditional physical measurements, which demonstrates the efficacy of the framework for accurate and fast variation analysis and statistical modeling.