The tomography reconstruction technique has widespread applications in arc diagnostics. This technology relies on arc radiation, with the reconstruction results representing the particle density at upper energy levels. However, tomography alone is insufficient to obtain the particle densities of other excited states and the electron density. To address this limitation, a tomography reconstruction algorithm combining a feed-forward neural network with a collisional radiative model (CRM) was proposed. The algorithm first trained the neural network using the CRM results, and then the tomography reconstruction data were fed into the network to predict the three-dimensional distributions of particle density and electron density. Experimental results demonstrate that the average prediction error is less than 5% when a two-hidden-layer feed-forward neural network is employed. Additionally, for the AMF vacuum arc, both the electron density and the particle density at lower levels are higher at the center of the arc and lower at the edges.