Application of 3D morphology and colour analysis for the assessment of acrylamide and hydroxymethylfurfural content in multi-grain biscuits

Thermal processing of foods high in carbohydrates presents safety concerns due to the formation of acrylamide and hydroxymethylfurfural (HMF). Traditional computer vision systems, which rely on two-dimensional (2D) colour proxies, lose accuracy in complex multi-grain matrices where surface browning and chemical changes often do not align. This research assessed a new multimodal framework that combines standard colour analysis with morphological descriptors obtained through monocular depth estimation using the "Depth Anything V2" algorithm. Biscuits made from six different flours (refined wheat, whole wheat, rye, oat, spelt, and buckwheat) were baked under various conditions. Levels of acrylamide and HMF were measured using LC-ESI-MS/MS and LC-DAD. A total of 35 features, including chromatic indices, 2D geometry, and three-dimensional topographical characteristics like quadratic roughness and mean curvature, were extracted from digital elevation models. Random Forest models significantly outperformed linear models, with global multimodal Random Forest models achieving a MAPE of 13.5% (R 2 = 0.93) for HMF and a MAPE of 7.7% (R 2 = 0.86) for acrylamide. Morphological descriptors proved to be crucial physical indicators of thermal stress in the crust model, capturing structural changes that traditional colourimetry cannot detect. This approach offers a non-destructive and cost-effective method for industrial compliance with Regulation (EU) 2017/2158 using affordable RGB imaging sensors.