DEEP-LEARNING STRATEGIES FOR FOOD IMAGE CLASSIFICATION INVOLVING FINE-TUNING AND DEEP FEATURE EXTRACTION

Abstract

With the exponential proliferation of food-related material on digital platforms, automatic food picture categorization has emerged as a critical study field. Deep learning models such as EfficientNetB0, Xception, and Inception-v3, which are known for their ability to use transfer learning, have become crucial tools in this sector. In this thesis, we critically evaluate the performance of such models on the complex Food-101 dataset that encompasses 101 various types of food. Our study found out that Xception is leading in its performance, with an awe inspiring accuracy rate of 84.54%, which surpasses other models. Based on this breakthrough, we explore how deep feature extraction techniques and powerful classification algorithms like SVM, Random Forest, and CatBoost can be integrated. Our findings prove how effective it is when we combine linear SVM with Xception attributes, which achieve a top accuracy of 93% for food image categorization. We have also analyzed the possibility of using features acquired from the pooling layer of EfficientNetB0 showing its superiority compared to others when linked to a Catboost classifier. This revolutionary study not only demonstrates the technological impact of these deep learning architectures but also shows their combined effects with machine learning classifiers, thereby advancing the frontier of accurate food image classification to new heights.