Résumé en anglais

A stable algorithm is proposed for image restoration based on the “mean curvature motion” equation. Existence and uniqueness of the “viscosity” solution of the equation are proved, a $L^\infty $ stable algorithm is given, experimental results are shown, and the subjacent vision model is compared with those introduced recently by several vision researchers. The algorithm presented appears to be the sharpest possible among the multiscale image smoothing methods preserving uniqueness and stability.