Comparing Machine Learning Based Segmentation Models on Jet Fire Radiation Zones

Risk assessment is relevant in any workplace, however there a degree of unpredictability when dealing with flammable or hazardous materials so that detection fire accidents by itself may not be enough. An example this the impingement jet fires, where heat fluxes flame could reach nearby equipment and dramatically increase probability domino effect catastrophic results. Because this, characterization such important from risk management point view. One would segmentation different radiation zones within flame, paper presents an exploratory research regarding several traditional computer vision Deep Learning approaches to solve specific problem. A data set propane fires used train evaluate given difference distribution background images, loss functions, seek alleviate imbalance, are also explored. Additionally, metrics correlated manual ranking performed experts make evaluation closely resembles expert's criteria. The Hausdorff Distance Adjusted Random Index were highest correlation best results obtained UNet architecture Weighted Cross-Entropy Loss. These can future extract more geometric information masks even implemented on other types accidents.