The main disadvantage of Magnetic Resonance Imaging (MRI) are its long scan times and, in consequence, its sensitivity to motion. Exploiting the complementary information from multiple receive coils, parallel imaging is able to recover images from under-sampled k-space data and to accelerate the measurement. Because parallel magnetic resonance imaging can be used to accelerate basically any imaging sequence it has many important applications. Parallel imaging brought a fundamental shift in image reconstruction: Image reconstruction changed from a simple direct Fourier transform to the solution of an ill-conditioned inverse problem. This work gives an overview of image reconstruction from the perspective of inverse problems. After introducing basic concepts such as regularization, discretization, and iterative reconstruction, advanced topics are discussed including algorithms for auto-calibration, the connection to approximation theory, and the combination with compressed sensing.