For duplicate image detection, the more advanced large-scale image retrieval systems in recent years have mainly used the Bag-of-Feature ( BoF ) model to meet the real-time. However, due to the lack of semantic information in the training process of the visual dictionary, BoF model cannot guarantee semantic similarity. Therefore, this paper proposes a duplicate image representation algorithm based on semi-supervised learning. This algorithm first generates semi-supervised hashes, and then maps the image local descriptors to binary codes based on semi-supervised learning. Finally, an image is represented by a frequency histogram of binary codes. Since the semantic information can be effectively introduced through the construction of the marker matrix and the classification matrix during the training process, semi-supervised learning can not only guarantee the metric similarity of the local descriptors, but also guarantee the semantic similarity. And the experimental results also show this algorithm has a better retrieval effect compared with traditional algorithms.