Deep hashing with mixed supervised losses for image search

Deep convolutional neural networks (DCNN) have revolutionized almost the whole computer vision fields, including learning to hash for image search. Recently, several supervised deep hashing methods are proposed to deal with large-scale image search, where most methods only consider one kind of supervised loss. In this paper, we show that image search performance can be further boosted by combining two kinds of supervised losses, by taking the combination of point-wise and triplet-wise losses as a study case. Two kinds of strategies are proposed to combine the strengths of them. One strategy is that the DCNN is first pre-trained with point-wise loss and then fine-tuned with triplet-wise loss. The other one is that the DCNN is trained jointly with point-wise and triplet-wise losses. We perform extensive experiments on two public benchmark datasets CIFAR-10 and NUS-WIDE. Experimental results demonstrate that the proposed methods outperform the compared methods with single supervised loss.