Pipeline for Calculating Calories for Print Recipes with Minimal User Intervention

Abstract

The thesis will provide a pipeline to estimate calorie counts from print recipes. The pipeline takes scanned recipes from cookbooks and uses Optical Character Recognition (OCR) to convert the scanned images of recipes to text. Several OCR tools were tested for their accuracy on fractions using a sample of the data, and the most accurate tool is used on the data. Next, a specially trained named entity recognition model is used to identify ingredients, quantities and units. These ingredients are used to search a database of values from the FDA to compute a calorie count for the recipe. The thesis tests the effectiveness of search by examining performance over 100 of the most common ingredients in the corpus of recipes. Finally, the thesis tests the performance of the model on a set of recipes, and found to estimate the calorie count at least as accurately as other automated approaches, such as those based on image recognition.