Deep Hashing Specialized for Multi-Label Image Retrieval

Abstract

In the era of big data, data hashing for approximate nearest neighbor(ANN) search is receiving a lot of attention. Data hashing enables fast retrieval and efficient storage in large data retrieval. Recently, deep learning-based hashing methods have improved the quality of retrieval by learning data feature extraction and hashing function learning at once. Most deep learning-based hashing methods use binary similarity to learn the similarity of data. In binary similarity, some pair of data has a similarity of 1 if any of them belong to the same class, and 0 otherwise. If the binary similarity is used for multi-label data, there is a limitation because priorities between similar data are not considered at all. In this thesis, the proposed model uses the proposed loss function and a novel evaluation metric, Mean RankedIOU(mRIOU), taking into account multi-label data. The proposed loss function handles multi-label data by giving a flexible boundary based on the Jaccard similarity. mRIOU evaluates performance based on the Jaccard similarity, enabling suitable performance evaluation for multi-label data. Experiments on multi-label datasets, MS COCO and NUS\_WIDE, indicate that the proposed method achieves better retrieval performance on both existing and proposed metrics. Qualitative evaluation is performed by visualizing the top 10 similar images for a given query image.