Flame Color Image Segmentation Based on Neural Network

A novel method of flame color image segmentation based on multilayer feedforward network is proposed. The training sample sets select the color and location information of the flame image in HSV color model as features. After preprocessing the training samples are normalized and input to multilayer feedforward network. By training with Levenberg-Marquardt algorithm the segmentation result is presented as a two-dimensional matrix which determines whether the pixel is a flame pixels or not with a suitable threshold. Experimental results show that the this method can segment flame image correctly, and is flexible to subsequent processing.