Relationships between colors and emotions have been studied for a long time in several domains, such as psychology and artistic theories. In this paper, we extract such relations appearing in social tagging systems, in which users can freely choose the images they upload and annotate, as well as the annotation tags. We first study two color representations that can be used to encode the chromatic contents of such images and select the most appropriate one for discovering color-emotion relationships, based on their performance for a classification task. We then extract, from this image corpus and based on the selected encoding, association rules characterizing relations between colors and emotions. We use the Apriori algorithm with a particular focus on the implications of color presence and absence on the emotion presences, commenting and discussing the obtained results.