This paper describes the flame vision system developed by CISE, on behalf of Thermical Research Division of ENEL, allowing a non-intrusive analysis and a probabilistic classification of the combustion process inside the gas turbines. The system is composed of a vision probe, designed for working in hostile environments and installed inside the combustion chamber, an optical element housing a videocamera, and a personal computer equipped with a frame grabber board. The main goal of the system is the flames classification in order to evaluate the occurrency of deviation from the optimal combustion conditions and to generate warning messages for power plant personnel. This is obtained by comparing some geometrical features (baricenter, inertia axes, area, orientation, etc.) extracted from flame area of images with templates found out during the training stage and classifying them in a probabilistic way by using a Bayesian algorithm. The vision system, now at the test stage, is intended to be a useful tool for combustion monitoring, has turbines set-up, periodic survey, and for collecting information concerning the burner efficiency and reliability; moreover the vision probe flexibility allows other applications as particle image velocimetry, spectral and thermal analysis.