A Color Quantization Optimization Approach for Image Representation Learning

Over the last two decades, hand-crafted feature extractors have been used in order to compose image representations. Recently, Representation Learning have been explored as a way of producing more intelligent features. In this work, we proposed two approaches of a Representation Learning method which aims to provide more effective and compact image representations. It uses Genetic Algorithm to improve the features of hand-crafted extractors by optimizing the colour quantization for the image domain. Our hypothesis is that changes in the quantization affect the description quality of the features enabling representation improvements. We evaluated the method performing experiments for the task of Content-based Image Retrieval on eight known datasets. The results showed that the first approach, focused on representation effectiveness, outperformed the baseline in all the tested scenarios. And the second, focused on compactness, was able to produce superior results maintaining or even reducing the dimensionality and representations until 25% smaller that presented statistically equivalent performance.