Deep Learning vs. Bag of Features in Machine Learning for Image Classification

The main issue in computer vision and notably image classification problems is image feature extraction and image encoding. Here we show and compare two approaches to solve this problem: the first approach uses the Bag of Features (BoF) paradigm. The second one is based on deep learning and especially Convolutional Neural Networks (CNN). Specifically, we use the "AlexNet" CNN model trained to perform well on the ImageNet dataset. Our results shed light on how the use of CNN is more performant than the BoF in the process of feature extraction in a machine learning framework for image classification. This performance is shown by a series of experimentations that we carried out using the Caltech dataset and many classifier algorithms.