Asymmetric hashing based on generative adversarial network

In the era of big data, social media, large-scale video, image, and text data is produced every day. The approximate nearest neighbor (ANN) search has drawn significant attention for content-based image retrieval applications to ensure retrieval quality and computational efficiency. Hashing has become a cutting-edge technology for image retrieval and big data applications due to its low-storage and high-computational efficiency. Hashing algorithms are useful for mapping images into short binary codes and generating a similar binary code for similar data points from the database. Many supervised/unsupervised hashing methods have been deployed for retrieving the query points from the database images, and many recently developed methods can achieve a higher accuracy regarding image retrieval performance. However, the current state-of-the-art algorithms can only improve binary code hashing, and the retrieval performance of binary representation is not good. To overcome this issue, we propose an asymmetric learning method that generates the hash codes. This work proposes a novel asymmetric learning-based generative adversarial network (AGAN) for image retrieval, which integrates the feature learning with hashing to an end-to-end learning framework. Moreover, to equip with the binary representation of image retrieval; we propose three loss functions, i.e., encoder loss, generator loss, and discriminator loss, which significantly improve retrieval performance. The extensive experiments show that our proposed method outperformed several state-of-the-art methods.