Assessment and ranking of 16 binarization methods for turbulent swirl flame images of various flame types

Abstract

Commercial digital cameras are versatile and low-cost non-invasive diagnostic tools for flame imaging with easy implementation for both industrial and laboratory flames. To quantify various flame metrics, flame image binarization is the first critical step in processing. Therefore, it is essential to know which binarization techniques perform well on different flame types, encompassing various operating conditions. This paper evaluates sixteen binarization methods for processing the chemiluminescent emission of eight natural gas and diesel oil flames captured by a commercial digital camera using three spectral filters at central wavelengths of 430, 515, and 590 nm. Proper binarization is challenging due to the blurry edges of the flame, which result from its finite thickness. If online data is required, processing time becomes critical, which is the first basis of the performance assessment of the sixteen binarization techniques. The second basis is a novel method for checking consistency, indicating the sensitivity of pixel assignment to foreground or background in a similar, yet new intensity neighborhood, which is critical for turbulent flames. The third metric was accuracy. It was defined as the fraction of the flame edge obtained from the binary image that was contained in the boundary region, determined based on the local standard deviations to avoid all biases associated with the use of a ground truth, making this basis novel and robust. The results were summarized for four different flame categories, concluding, for example, that the widely used Otsu method performs well in all cases.