Comparative analysis of image segmentation methods on various flame types and their influence on flame stability assessment

Image-based flame diagnostics are crucial in optimizing high-temperature processes, improving efficiency, and minimizing harmful emissions. A pivotal step in this process is flame segmentation. Despite its importance, there is a lack of studies examining the segmentation of low-brightness flames. In many industrial scenarios, flame images suffer from challenges such as blurry boundaries and inadequate brightness. Therefore, in this study, different segmentation techniques were explored. To assess the robustness of each method, flames from three distinct feedstocks were optically observed within a multi-feed test facility. Various techniques were evaluated, including Otsu's method, manual thresholding, multilevel thresholding, k-nearest neighbors (KNN), and deep learning (U-Net). Then, geometrical and positional flame characteristics were derived. It was found that, in low-brightness flame scenarios, classic methods tend to produce inaccurate results, with a segmentation quality of 60–67 %. Furthermore, additional image preprocessing steps are necessary to effectively segment such flames to enhance the flame's appearance. U-Net emerged as the most promising among the tested methods, achieving a segmentation quality of around 89 % for low-brightness flames. Although it provided nearly accurate flame shapes, aligning well with the ground truth. Furthermore, a new approach was introduced to estimate global flame stability by considering factors such as flame orientation, ignition point, and geometric parameters, thereby enabling a comprehensive stability assessment. It was observed that solid fuel flames exhibited the highest stability index, around 98 %. Additionally, it was evident that inaccuracies in segmentation quality significantly impacted the stability index, highlighting the importance of robust segmentation steps in flame stability assessment.