The traditional flame detection algorithm often achieves incomplete contour and poor anti-interference performance in the process of flame foreground extraction. This paper proposes a new flame foreground extraction algorithm, which combines RGB, HSI, and Ostu (maximum between-cluster variance method). The developed algorithm can extract flame contour completely and eliminate the smallest possible interference. Then, static features such as textures and colors in YCbCr are extracted by using a co-occurrence matrix and used for final flame judgment. Finally, an improved probabilistic neural network (PNN) method is developed to adjust the traditional smoothing factor from a single fixed value to a parameter that contains multi-variables, after which the expectation/conditional maximization (ECM) algorithm is used to find the optimal parameters. The extracted features are input in the advanced PNN and used for the training test. Simulation results show that the proposed algorithm can improve the accuracy of flame identification with good anti-interference performance.