TinyML-Based Bean Cooking Time Prediction Using Cooking Data and Bean Images

The cooking time variability of legumes, specifically the common bean (Phaseolus vulgaris L.), poses a major challenge for the food industry and food security in Africa. Traditional estimation methods are destructive and time-consuming, making them unsuitable for rapid field testing. This paper presents a non-destructive approach based on Tiny Machine Learning (TinyML) to predict cooking time using RGB imaging.We compiled a regression-specific dataset (Dreg) comprising 11,200 images annotated with continuous cooking times ranging from 51 to 410 minutes. We compare the performance of transfer learning architectures (MobileNetV2, EfficientNet-B0) against custom convolutional networks, TBNet2 and TBNet5, specifically designed for constrained environments.The results demonstrate the superiority of TBNet2 for the regression task, achieving a Mean Absolute Error (MAE) of 16.40 minutes and an $R^2$ coefficient of 0.90, in contrast to the divergence observed in pre-trained models. Furthermore, a half-precision (Float16) quantization strategy allows for the deployment of this model on a smartphone with a latency of less than 160 ms, validating the feasibility of an embedded and autonomous diagnostic tool.