For signal processing related to localization technologies, non line of sight (NLOS) multipaths have great impact over the localization error level. This study proposes a localization correction method based on convolution neural network (CNN) that extracts obstacles' features from maps to predict the localization errors caused by NLOS effects. A novel compensation scheme is developed and structured around the localization error predicted by the CNN. Four prediction tasks are executed over the building distributions within the maps and the propagation model in urban zones, resulting in CNN models with high prediction accuracy. Finally, a thorough comparison of the accuracy performance between the time difference of arrival (TDOA) localization algorithm and the results after the error compensation reveals that, generally, the CNN prediction approach demonstrates a great localization error correction performance. It can be observed that the powerful feature extraction capability of CNN can be exploited by processing surrounding maps to predict localization error distribution, which has great potential in further enhancement of TDOA performance under challenging scenarios with rich multi-path propagation.