基于多神经网络分类器组合的火焰图像分割

Multiple Neural Networks Classifier Combination for Segmentation of Flame Images

单一神经网络分类器的性能很大程度上取决于网络参数的选择 ,设计一个性能最优的神经网络分类器是非常困难的。针对这一问题 ,本文提出了基于多个 BP神经网络分类器组合的回转窑火焰图像分割方法。选取多组不同的训练样本对多个具有不同初始条件的 BP网络进行训练 ,网络收敛后 ,用于火焰图像的分割 ,会产生多种分割结果 ,采用平均值法、投票表决法、最大统计概率法和神经网络 4种方法对其进行组合 ,得到了最终的分割结果。实验结果表明 ,本文提出的方法具有分割效果好和可靠性高等优点 ,满足了实际使用的要求。

The performance of a single neural network classifier depends on the selection of parameters of the neural network to a great extent, so the design of a neural network classifier with the best performance is very difficult. In this paper, a new approach to segmenting flame images in kiln based on multiple BP network classifier combination is proposed. First, this paper selects some different samples to train multiple BP neural networks with different initial values, then the converged BP networks are used to segment flame images and finally, the obtained results are combined by four algorithms, i.e., mean, majority voting, maximum statistical probability and neural network, to get the final result. Experimental results demonstrate that the method presented here can achieve high segmentation accuracy and high reliability, which suffices the requirement of practical use.