Noisy-label-adaptive SegFormer parameter optimization for high-enthalpy flow PLIF image segmentation

Semantic segmentation of optical images achieves precise delineation of target contours and spatial distribution characteristics through pixel-level classification. This technique serves as a vital tool for combustion diagnostics and flame characterization, enabling quantitative analysis of flame structure, dynamic evolution, and physicochemical properties. However, PLIF (Planar Laser-Induced Fluorescence) image segmentation in high-enthalpy flow fields faces significant challenges, including turbulent interference, thermal quenching effects, background noise, and low signal-to-noise ratio, which result in blurred flame boundaries and non-uniform intensity distribution, severely compromising the accuracy of conventional segmentation methods. Although neural networks can improve segmentation precision, their reliance on large-scale annotated datasets remains problematic–data acquisition is costly and manual labeling introduces subjective variability. To address these limitations, we propose a novel approach that treats labels generated by conventional methods as noisy label data. By integrating parameter fine-tuning techniques with the SegFormer architecture, our method demonstrates significantly enhanced noise robustness. Experimental results confirm substantial improvement in PLIF image segmentation accuracy while dramatically reducing annotation requirements, providing a practical solution for advanced combustion analysis.