Deep Feature Consolidation for Perceptual ImageHashing: A Pre-Trained CNN Approach

Perceptual image hashing is utilized in various applications, including image search, image manipulation detection, image copyright violation detection, fingerprint matching, and zero watermarking, among others. Conventional image hashing methods extract various local, transformed, statistical, or latent features and quantize them to consolidate into vec- tors that serve as hashes. Learning-based methods utilize deep neural networks to generate high-quality image hashes. In this paper, we present a learning-based image hashing mechanism that consolidates a large num- ber of low-level image features obtained using a stack of pre-trained con- volutional layers from multiple convolutional neural networks to construct image hashes from higher-level features with minimal training. We also offer both graphical and empirical analyses of our mechanism, including ablation analysis and comparisons with state-of-the-art algorithms.