Abstract: Accurate identification of the tobacco leaf curing stage is crucial for ensuring the final product quality. However, traditional manual methods suffer from subjectivity and delayed response. To address the issue of insufficient generalization capability of existing models in complex environments, this paper proposes TriLight-Ensemble—a lightweight model ensemble framework based on HSV color space conversion. Firstly, the feature representation capability is enhanced through RGB-to-HSV color space conversion. Secondly, a progressive transfer learning strategy is adopted to optimize some convolutional layers of three lightweight networks (MobileNet v2, EfficientNet b0, and ShuffleNet v2). Finally, the ensemble method based on a hard voting mechanism coordinates the decision-making of multiple models. Experiments on the tobacco leaf curing stage dataset (224×224 pixels) demonstrate that TriLight-Ensemble achieves a recognition accuracy of 91.03%, representing a 3.16% improvement over the ShuffleNet v2 baseline model. This method not only provides a reliable solution for intelligent curing monitoring but can also be extended to other agricultural product state recognition fields (such as tea processing or grain drying).