Real Time Deep Learning Model for Food Item Identification and Recipe Data Generation

Accurate and fast identification of various food items can be very useful, in terms of preventing harm caused by allergies and other problems. In this work, an attempt is made to classify food items from real-world images with great accuracy and in real time. The proposed model is designed in a dual-phase classification method incorporating unsupervised clustering followed by a classification model to improve the initial prediction results. The dataset, selected primarily for its various food classes, consists of real-world images of food, captured outside controlled conditions. During the first stage, clustering is used to group class predictions into clusters. In the next stage, the model for the cluster with the highest prediction value is loaded to make the final prediction. Each cluster model is trained on a small subset of classes, reducing time and cost thereby improving performance. The accuracy metrics of the general model and some of the sub-models are compared to see if using smaller label subsets provides improved performance without a large increase in training time. Finally, for the generation of detailed information about food items and suggested recipes, an LLM will be integrated into the proposed model. Custom prompts will be used to generate contextually relevant data more effectively