Machine Learning-Augmented Flame Front Detection in High-Speed Imaging: A Physics-Guided Framework

High-speed combustion imaging provides critical insight into ignition, flame-kernel evolution, and turbulent flame propagation in advanced propulsion systems. However, extracting physically meaningful flame boundaries from such data remains challenging due to low signal-to-noise ratio, rapid luminosity variation, window fouling, and interference from shot noise and internal chemiluminescence gradients. Conventional image-processing workflows, typically based on thresholding, morphological operations, or classical gradient filters, require extensive manual tuning and often produce fragmented or non-physical flame contours that do not generalize across operating conditions. This work presents AGNI AI-Guided aNalysis for Image data. We currently evaluate the AGNI framework for flame-edge detection using high-speed OH∗ chemiluminescence images from a spark-ignited hydrogen-fueled heavy-duty optical engine. AGNI is modular, and automated. We achieve this by using a hierarchical classification approach that integrates three components: a usability classifier that removes non-informative images such as background data, and quenched flames, a segmentation model for extracting the flame boundary and a third artifact removal module that eliminates shot noise, double flame edges and other aberrations to produce a physically consistent flame front. We evaluate the model on multiple cycles of high-speed OH* chemiluminescence images. We observe that the model produces consistent high-quality flame front data under multiple operating conditions. The results demonstrate that the proposed method significantly outperforms traditional ImageJ and MATLAB workflows by suppressing noise-induced false detections, preserving flame topology, and maintaining robustness under severe synthetic perturbations, including high-intensity Gaussian noise and frame-level signal degradation. Additionally, unlike traditional techniques that require manual intervention for post-processing and alignment, AGNI requires minimal supervision and can produce high quality flame fronts with limited training data.