Cellpose-SAM: Transformer-based segmentation for cellular instability analysis of spherical hydrogen-air premixed flames

• Introduced Cellpose-SAM for Transformer-based flame segmentation. • Achieved accurate results with fewer than 20 annotated images. • Ensured physically consistent extraction of flame morphology metrics. • Linked cellular instability to macroscopic flame acceleration. • Outperformed U-Net in contour fidelity and cell detection accuracy. Accurate identification of cellular structures in hydrogen–air premixed flames is essential for understanding cellular instabilities and acceleration mechanisms. Traditional image processing and convolutional neural network-based models often fail under highly wrinkled or topology-varying flame fronts. This paper proposes Cellpose-SAM, a Transformer-based segmentation framework that integrates the Segment Anything Model into the Cellpose pipeline to achieve foundation-model-guided feature recognition. Schlieren experiments were performed in a constant-volume vessel at equivalence ratios ( ER ) of 0.8, 1.0, and 1.5. The model exhibits rapid and stable convergence, achieving its minimum loss within 60 epochs and maintaining consistent performance with fewer than 20 annotated samples. The extracted morphological parameters, including cell count, contour length, and average cell area, show clear regime-dependent evolution, with lean flames 0.8 displaying the highest cell density and wrinkling intensity. Derived physical metrics such as flame speed, wrinkling factor, and cell-line density confirm a self-consistent acceleration mechanism, where enhanced front distortion leads to increased propagation rate and higher power-law exponent a from 0.96 to 1.10. Comparative analysis further demonstrates that Cellpose-SAM provides finer contour continuity and more physically accurate segmentation than traditional UNet-based methods, ensuring accurate extraction of dynamic instability features. A visualization comparison of the segmentation and flame evolution process is provided in the supplementary material ( https://reurl.cc/4lpV7v ).