Abstract: Edge detection based on Mamdani fuzzy inference regards all features as a whole to carry on one-step inference. However, this will decrease the robustness of algorithm since different features are difficult to guarantee both the sensibility for fuzzy edge and the suppression for noise. This paper presents a step-by-step processing scheme based on multi-features importance-weighted fuzzy inference. According to the importance degree of each feature in the edge representation and noise suppression, the proposed method carries on a step-by-step fuzzy inference and weights all inference results to obtain the edge membership. In addition, this paper presents an improved centroid method based on area approximation, which has higher precision and real-time in defuzzification. The experiment results show the precision, robustness and real-time of the proposed algorithm.