Abstract
This paper introduces TD hashing, a novel image hashing technique that utilizes Tucker decomposition to create compact, secure, and highly robust digital signatures. By treating image data as a three-order tensor rather than a flat matrix, the method achieves a superior balance between robustness and discrimination. Experimental results demonstrate that TD hashing outperforms several state-of-the-art algorithms, achieving a near-perfect Area Under the Curve (AUC) of 0.9993.
Introduction
In large-scale image retrieval and authentication, digital hashes serve as short binary representations of image content. An effective image hash must balance two competing properties: robustness and discrimination. Robustness ensures that visually identical images (e.g., compressed versions) produce nearly identical hashes, while discrimination ensures that distinct images produce significantly different hashes.
Most existing methods, such as those based on histograms or standard matrix factorizations like Singular Value Decomposition (SVD), struggle with this balance, particularly when images undergo geometric transformations like rotation or cropping. These methods typically treat images as flat 2D matrices, losing the intrinsic multi-dimensional structure of the data. The authors propose that tensor-based decomposition can better preserve this structural information.
Methodology
The proposed TD hashing pipeline consists of four primary stages: preprocessing, tensor construction, Tucker decomposition, and hash generation.
Preprocessing and Tensor Construction
Images are first resized and converted to the CIE $L^*a^*b^*$ color space. The $L^*$ (lightness) component is selected for its alignment with human visual perception. A Gaussian low-pass filter is then applied to mitigate the effects of minor noise.
To ensure security, the image is divided into blocks which are randomly selected and stacked into a three-order tensor (a 3D data cube) using a secret key. This key-dependency ensures that an attacker cannot easily forge a hash without the secret key.
Tucker Decomposition and Hash Generation
The core innovation is the application of Tucker decomposition to the 3D tensor. This process yields a "core tensor" and three orthogonal factor matrices that act as a compressed summary of the image's intrinsic structure. The rows of these factor matrices are averaged and quantized to produce a compact 96-bit binary sequence. Similarity between images is then measured using the Hamming distance between their respective hashes.
Experimental Evaluation
The researchers evaluated their method using 14,551 images across three benchmarks: the Kodak Lossless True Color Image Suite, INRIA Copydays, and the UCID database. To simulate real-world conditions, they applied 72 different manipulations, including JPEG compression, speckle noise, scaling, and complex "rotation-cropping-rescaling" attacks.
Key Results
- Superior Accuracy: TD hashing achieved an AUC of 0.9993, outperforming competitors like GF-LVQ (0.9929) and SVD-CSLBP (0.9601).
- Efficiency: The average time to extract a hash is approximately 0.17 seconds, making it fast enough for practical use.
- Compactness: At only 96 bits, the hash is significantly smaller than many other robust methods, which can require 720 bits or more, making it ideal for large-scale database indexing.
Discussion and Conclusion
The stability of the factor matrices produced by Tucker decomposition is central to the method's success. By capturing spatial correlations across the image more effectively than 2D methods, TD hashing maintains structural consistency even under distortion. Statistical modeling using Chi-square tests confirmed that the Hamming distances follow a Weibull distribution, providing a theoretical foundation for setting optimal similarity thresholds.
While highly effective, geometric attacks involving simultaneous rotation, cropping, and rescaling remain the most significant challenge. Future work will explore extending this framework to video hashing and investigating how deep learning features could be integrated into the tensor construction phase to further improve robustness.
Summary updated