Variational autoencoders are commonly used for image compression, reconstruction, and generation. This thesis studies the integration of a pre-processing step, based on the Discrete Cosine Transform (DCT), within classical and quantum variational autoencoders. The main objective is to assess the possible improvements that can be obtained by applying a frequency-based dimensionality reduction before the encoding stage. The main idea behind the choice of the DCT as pre-processing technique is due to the sparse nature of the MNIST images after the transformation has been applied. In fact, for both classical and quantum per-processing, the information is mostly concentrated in the same features, usually corresponding to the lower frequencies.