Pseudo pairs based unsupervised deep hashing for image retrieval

Abstract

The research presents a new deep learning framework, the pseudo-pair-based unsupervised deep hashing (PPUDH), designed to enhance image retrieval systems. PPUDH employs a soft clustering approach that iteratively trains clusters with strong discriminative capabilities and creates binary codes (BCs) with heightened correlation sensitivity. These clusters are then amalgamated to form an additional distribution for deriving hash codes (HCs). The model undergoes optimization via standard stochastic gradient descent (SGD). This optimization process marries the reconstruction loss from the encoder tasked with auto-reconstruction with the loss incurred from meeting binary code requirements. The efficacy of PPUDH has been validated through comprehensive evaluations of three renowned datasets. The outcomes of these tests demonstrate that PPUDH offers a considerable advancement over existing top-tier methods in the field.