In image classification, most of the highest-performing statistical learning approaches are based on the Bag-of-Words model. In this PhD thesis, we propose an extension of this formalism. Our aim is to advance the state-of-the-art by introducing a density function-based pooling strategy. Our proposed BossaNova representation is based upon that novel pooling strategy, enhancing the Bag-of-Words model. The experimental evaluations on many challenging image classification benchmarks showed the advantage of BossaNova when compared to traditional techniques. Moreover, our participation in the ImageCLEF 2012 challenge achieved the 2nd rank among the 28 visual submissions. BossaNova has the potential to become a new standard representation in image classification tasks.