Perceptual Image Hashing with Bidirectional Generative Adversarial Networks for Copy Detection

Abstract

The traditional perceptual hashing algorithm is generally achieved by artificially extracting features from an image and quantizing them into a hash code. However, it is often hard to capture the inherent features of an image in practice. In order to improve the performance of image perceptual hashing, an unsupervised data-driven generative adversarial framework to generate the image perceptual hash code is proposed in this paper. Firstly, the original image is normalized and the encoder network is employed to generate the perceptual hash code; then, the generator network is used to formulate a data distribution as similar as possible to the original image from random noise in the same dimension as the perceptual hash code. Thirdly, the discriminator network is adopted to distinguish the hash code from the noise modified by the generated network and the source of the generated image and the original image. Finally, the encoder can generate image hash codes with good perceptual robustness and recognition accuracy by jointly training of the three networks. Various experiments have been executed on an extensive test database in this paper. The results show that the proposed perceptual image hashing algorithm has stronger robustness than other state-of-the-art schemes.