This article presents a novel technique for automatic archaeological sherd classification. Sherds that are found in the field usually have little to no visible textual information such as symbols, graphs, or marks on them. This makes manual classification an extremely difficult and time-consuming task for conservators and archaeologists. For a bunch of sherds found in the field, an expert identifies different classes and indicates at least one representative sherd for each class (training sample). The proposed technique uses the representative sherds in order to correctly classify the remaining sherds. For each sherd, local features based on color and texture information are extracted and are then transformed into a global vector that describes the whole sherd image, using a new bag of words technique. Finally, a feature selection algorithm is applied that locates features with high discriminative power. Extensive experiments were performed in order to verify the effectiveness of the proposed technique and show very promising results.