COOKING RECIPE GENERATION USING YOLOV11 AND T5V1.1

Abstract

Eating is an essential aspect of human life, yet deciding what dish to prepare based on available ingredients can be a hassle. This paper presents an approach to automated recipe generation by integrating a multimodal system consists of an ingredient detection model based on YOLOv11 [1] with a recipe generation model utilizing T5v1.1 [2]. The system processes an input image of available ingredients to generate a recipe derived from the detected components. We trained the YOLOv11 model on the AICook [3] dataset, enabling it to accurately detect and recognize ingredients from images. For recipe generation, we fine-tuned the T5v1.1 base model on a modified version of the RecipeNLG [4] dataset, which consists of approximately 2 million recipes. The proposed system achieved a mean average precision (mAP) of 0.971 at IoU@50 for ingredient detection and BLEU scores of 0.667, 0.408, 0.302, and 0.24 for BLEU-1, BLEU-2, BLEU-3, and BLEU-4, respectively, in recipe generation. These results demonstrate the system's potential to simplify meal preparation by automating the recipe generation process.