Sexual exploitation of children is without a doubt one of the most heinous crimes in modern society, prompting us to the need of responding firmly with the development and implementation of strict laws as well as the development of efficient and effective means to detect such activities. In this article, we propose an initial solution to the problem of child pornography detection in digital images aiming at assisting lawenforcement agents on investigating criminal activities related to the generation and distribution of this type of content. Our solution relies upon a two-tier representation of image features calculated through the concept of visual dictionaries with distance-preserving properties of low-level features enabled by the recently proposed BossaNova formalism. We validate it with thousands of images of real-world apprehensions of the Brazilian Federal Police. The results, although still far from solving the problem, show that this solution is promising.