Quaternion algebra allows to generate new representations of images which capture non-trivial equivariant features related to illumination changes and rotations. We compare the performance of this feature representation against a conventional CNNs layer using four datasets. The most important result is that the accuracy gained by using this representation is substantially more robust to illumination changes than nets without such a layer. In addition to present recent results, we will also further discuss current and future work on possible deep learning applications of this representation.