Bitumen extraction from oil sands is carried out in large separation cells using a water-based extraction process. Controlling the Bitumen-froth and Middlings interface in these cells at an optimum level provides significant economical and environmental benefits. Traditional sensors are not reliable in estimating this interface level and novel vision-based sensors have been developed previously to overcome this problem. These sensors estimate the interface level and its confidence for separation cells with a single sight glass. The confidence value only represents the turbidity of the interface and hence cannot be used for control decision making in all process conditions. The current work describes an image differencing algorithm for interface level detection which also facilitates the computation of a confidence estimate that is accurate in most process situations. The confidence value is computed based on noise statistics, an appropriately chosen edge detection method and a change detection algorithm. Another significant advantage of the algorithm is that both the interface level and the confidence estimation procedures can be extended in a straight-forward manner to handle the presence of multiple sight glasses. Offline results show that the algorithm accurately detects the interface level in normal process conditions (with high confidence values) and outputs correct confidence values in other situations, with very low false-positive and false-negative error rates. Online industrial implementation results show that the vision sensor tracks the interface level very closely and results in significant automation of plant. © Springer-Verlag 2011.