Deep Learning Based Hybrid Framework for Food Classification and Recipe Generation

In this paper, an attempt is made to improve the classification of food items from real-world images in addition to images sourced from controlled conditions. The dataset, which was selected for its large number of food classes, consists of 101,000 real-world food images divided equally between 101 class labels. The proposed model is designed in a two-stage methodology using a general model trained on all classes to obtain a heuristic value followed by use of a specialized sub-model for fine-tuned classification. In the first stage, a general model groups and adds up the predictions by cluster, providing a confidence value for each cluster. In the second stage, the model for the cluster with the highest confidence value is used for the final classification. Each cluster model is trained on a small subset of classes, reducing time and cost along with improving the performance. The study evaluates whether using smaller label subsets in sub-models leads to better classification performance in comparison to the general model while keeping training time manageable. The sub-models reach over 95 percent accuracy after training for 30 epochs, while the general model achieves 82 percent after 50 epochs. Therefore, the sub-models outperform the general model on the specific label subsets they were trained on, and take converge faster.