A multispectral camera setup is used to infer a 2D map of plasma parameters in a tokamak from spectral emissions. However, the light measured by these cameras is line integrated in the toroidal direction, whereas emissivities on the poloidal plane are necessary for the inference. The poloidal plasma emissivity can be obtained by tomographic reconstruction, but classical techniques are too slow to use these emissivities for real-time control. We present two machine learning based approaches to accelerate the reconstruction of the poloidal emissivities from line integrated data measured by the camera setup. Both approaches yield more accurate results on synthetic data than the iterative approach while being, with the right implementation, fast enough for real-time control applications.