Accurately detecting multiple simultaneous touches is crucial for various applications using piezoresistance sensor arrays. However, calibrating them is difficult due to their nonlinearity and hysteresis. While data-driven deep learning approaches could model complex sensor patterns, the required amount of labeled data increases exponentially as the number of contact points or sensor subelements increases. In this letter, we propose a novel supervised learning framework, Local Message Passing Network, that only needs single touch data to calibrate multiple contact points into a high resolution pressure map. The individual sub-local networks eliminate domain shift problems, while a message passing mechanism enables them to correctly learn correlations between neighboring sensor subelements. The performances of the proposed model were tested on labeled single- and double-pressure data and compared with previous deep learning calibration methods. Experimental results show that our framework can expand prior knowledge of single touch data to calibrate multi-touch sensor inputs into high resolution pressure maps.