A new approach of graph cuts based segmentation for thermal IR image analysis

Abstract

Thermal Infra Red images are one of the most investigated and popular data modalities whose usage has grown exponentially from humble origins to being one of the most extensively harnessed imaging forms. Instead of capturing the radiometry in visible spectra, the thermal Images focus on the near to mid Infrared spectra thereby producing a scene structure quite different from their visual counterpart images. Also traditionally the spatial resolution of the infra red images has been typically lower than traditional color images. The above reasons have contributed to the past trend of minimal automated analysis of thermal images wherein intensity (which corresponds to heat content) and to a lesser extent spatiality formed the primary features of interest in an IR image. In this work we extend the automated processing of Infra red images by using an advanced image analysis technique called Graph cuts. Graph cuts have the unique property of providing global optimal segmentation which has contributed to its popularity. We side step the extensive computational requirements of a Graph cuts procedure (which consider pixels as the vertices of graphs) by performing preprocessing by performing initial segmentation to obtain a short list of candidate regions. Features extracted on the candidate regions are then used as an input for the graph cut procedure. Appropriate energy functions are used to combine traditionally used graph cuts feature like intensity feature with new salient features like gradients. The results show the effectiveness of using the above technique for automated processing of thermal infrared images especially when compared with traditional techniques like intensity thresholding.