Image classification is a subject with a rich history. Most approaches can be separated into two distinct sub-classes of problems, feature extraction from images and multiclass classification in a high dimensional feature space. In this project we provide an empirical analysis of three approaches to this problem.
We first evaluate the performance of a k-Nearest Neighbour classifier using features derived from a low-resolution image representation (’tiny image’). We show that the performance of this classifier can be superseded by an ensemble of 15 one-vs-all classifiers trained on a visual bag of words lexicon derived from raw image pixels.
Finally, we show that a multi-class logistic regression classifier coupled with features extracted by spatial pyramid matching outperforms each of the aforementioned methods.
More Details can be found in our Report.