Devised Deephashing Models Through Self-Supervised Learning for Image Retrieval

Many image database systems utilize content-based retrieval technology using perceptual hashing to facilitate image comparison and retrieval. However, conventional hashing functions generate different hash codes for the same image domain even with minor changes, making accurate searches in the image database impossible. To overcome these limitations, a method has been devised to enable the search for images similar to queries synthesizing visual feature information, including texture, shape, and color of images. This paper proposes deephashing models that incorporate both variational autoencoders and vision trans-formers, capable of capturing semantic patterns and preserving the semantic similarity between data points. To validate this proposal, we conduct a comparative evaluation by comparing its performance with widely-used supervised learning-based models. The experimental analysis shows that the proposed approach generates compressed and discriminative hash codes with fixed-length binary representations that maintain semantic similarity and enable the search even for similar images. When evaluating the performance of our proposed model in comparison to the other studies, noticeable improvements are observed across the selected metrics in deephashing based on self-supervised learning.