In magnetic particle imaging, many applications require the time consuming measurement of a system matrix before image reconstruction. Reduction of measurement time can be achieved with the help of compressed sensing, which is based on the sparsity of the system matrix in a suitable transform domain. In this work, we propose regularization functions to exploit the additional correlations in multi-patch system matrices. Experiments show that the resulting recovery method allows successful matrix recovery at higher undersampling factors than a standard compressed sensing recovery.