On an image segmentation method for forest fire metrology

The improvement of forest fire metrology tools constitutes a crucial step towards firefighting means optimization. In this context, the aim of this work is to extract accurately fire regions in images. Chrominance channels of Lab, YUV and YCbCr color spaces are investigated to find a discriminative channel for fire pixels segmentation. Based on an optimized k-means clustering in the Cb-channel of YCbCr and on distribution properties of fire pixels in the RGB color space, a new method is proposed. The performances of this method are first qualitatively presented. A quantitative evaluation is then conducted using three supervised evaluation criteria based on expert annotations. Finally, a comparative study is performed with other existing algorithms assessing the good behavior of the proposed image processing method.