We propose a deep learning based measurement model for Monte Carlo self-localization to improve the robustness to robot calibration errors in the RoboCup Humanoid League context. A deep neural network estimates the distance between a line mask computed by the vision pipeline from a camera image and given pose candidates. This distance is then used to derive weights for the particles of the Monte Carlo localization. The network is trained using images generated by applying a perspective transformation on the map of the field. We use a style transfer model to preprocess the input data for finetuning the model to build robustness towards noise and detection errors in real-world data. Our model is more robust to erroneous head pose measurements than the baseline. Further, we show significant improvements in global localization.