Deep Hash Learning of Feature-Invariant Representation for Single-Label and Multi-label Retrieval

In large-scale retrieval, hash learning is favored by people owing to its fast speed. Nowadays, many hashing methods based on deep learning are proposed, because they have better performance than traditional feature representation methods. Both in supervised hash learning and unsupervised hash learning, similarity matrix is used in the objective function. In the similarity matrix, if two images share at least one label, the similarity is "1", otherwise it is "0". However, this kind of similarity can not reflect the similarity ranking of multi-label images well, which is vulnerable to pixel interference. Therefore, in order to improve the retrieval accuracy of multi-label data, we improve the traditional deep learning hashing method by dividing the multi-label images into "strong similarity" and "weak similarity". In addition, although the deep neural network can judge the label of the images directly through the pixels, it does not understand the high-level semantics of the images. Hence, we take the feature invariance of the images into consideration, which means that the transformed image should have the same feature representation with the original image. In this way, we propose a novel Deep Hash learning method based on Feature-Invariant representation (FIDH), which focuses on deep understanding rather than deep learning. Experiments on common single-label and multi-label datasets show that our method obtains better performance than state-of-the-art methods in large-scale image retrieval.