EfficientNetB0-CatBoost: Deep Learning With Categorical Boosting for Food Image Recognition

Research into machine learning and deep learning methods for image categorization has been sparked by the explosion of visual content available on digital platforms, with particular interest in domains such as food image recognition. With so many food-related images available on the internet and social media platforms, automated food image categorization is a major difficulty. Deep Convolutional Neural Network architectures offer a promising approach by leveraging pre-trained models and transfer learning to extract meaningful image representations. EfficientNetB0 is an example of a pre-trained deep learning architecture that has demonstrated impressive ability in extracting high-level features from images. This study explores the effectiveness of using deep features extracted from the average pooling layer of a fine-tuned EfficientNetB0 model for the classification of complex food images using the Food-101 dataset. Subsequently, the features extracted are passed as input to a Categorical Boosting (CatBoost) classifier. To assess performance, we compare the classification accuracies of CatBoost, Support Vector Machines with various kernels, and a Random Forest classifier, all utilizing the same set of extracted features, against the accuracy of the end-to-end fine-tuned EfficientNetB0 model. Our experiments reveal that the proposed EfficientNetB0-CatBoost framework achieves the highest classification accuracy for classifying complex food images, highlighting the potential of CatBoost for efficient classification of deep features.