Learning invariant feature toward recognition modeling biology vision

Generalizing object recognition to be invariant to geometric transformation is a traditional challenge in vision. Inspired by information processing mechanism in biology vision, we develop a invariant recognition model by exploiting temporal correlation. Maximizing spatial independence leads to emergence of simple cell properties. Subsequently minimizing the variation of simple cells outputs over time leads to emergence of invariant features typical of complex cells. Experiments on character images testify recognition is rotation and translation invariant. Our model's plausibility in neurobiology view is also discussed.