Variations in run-of-mine ore properties such as size, composition, and grindability strongly affect AG and SAG mills performance. In the past, most efforts to track and control these variations have focused on developing on-line size analysis using vision systems to reduce power consumption and avoid mill overload. To improve ore sorting/blending strategies and to account for variations in grindability, on-line run-of-mine composition analysis would be very helpful, but has received much less attention from the research community. This paper describes a general machine vision approach for on-line estimation of rock mixture composition, and is illustrated on a very challenging nickel mineral system: very heterogeneous minerals, similar coloration, and rock fragments can be dry or wet. The proposed mineral type recognition method involves: (1) dividing into sub-images; (2) extracting color and textural features using principal components analysis (PCA) and wavelet texture analysis (WTA), respectively; (3) reducing feature space dimensionality and removing dry/wet systematic variations using discriminant partial least squares (PLS-DA); and (4) establishing class boundaries using support vector machines (SVM). Through a pilot plant conveyor belt application, very good results were obtained for dry minerals. For wet rock mixtures, further investigation is required, but results are very promising. (c) 2007 Elsevier Ltd. All rights reserved.