Abstract: A new design method based on multilevel SVM classifier is presented to recognizethe normal sintering, oversintering and undersintering states in rotary kiln. The ROIs (regions of interesting) were segmented from the images of sintered materials by an improved dual-fast marching method, including the zones of materials, blackbar, flame and full calcinations. Then, the characteristics of colour, shape and vein were extracted from ROIs to develop a pre-treated classifier model based on One-Versus-Another method to classify the different sintering states. The distribution of wrong sample points provided by the pre-treated classifier was studied, and the sample points that are easy to confuse with each other were classified as the same class. Based on the multilevel SVM, a new classifier model was thus redeveloped for different sintering states. The experimental results showed that the new model is fast and accurate with wide applications.