Deep Hashing Capable of Adding New Dataset without Class Labels

Deep hashing has realized very accurate similar image retrieval by learning the hash function via deep neural networks (DNN). Most deep hashing methods assume that queries originate from the same image dataset as the training images. However in practice, as the search system becomes older, it will encounter more query images that are not covered by the learned image dataset, To cope with this issue, we propose a solution named UIDH (Unsupervised Incremental Deep Hashing) that enables a deep hashing trained on one image dataset to learn an additional image dataset as new training data. First, UIDH obtains the initial hash function that outputs
$K$
-bit hash codes by learning the first image dataset with supervision. Later, UIDH learns the additional image dataset with unsupervised incremental learning and derives the second hash function that outputs
$\frac{K}{2}$
-bit hash codes. The final hash function in UIDH concatenates the upper half of the initial function with the second hash function. It maintains the high retrieval accuracy for the first dataset thanks to the initial supervised hashing and improves the similarity search for the additional dataset supported by the second unsupervised hashing. In this way, UIDH realizes an image search system that can treat the two dataset simultaneously. The novelty in UIDH is to handle the additional dataset without class labels.