Robust image hashing for content identification through contrastive self-supervised learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Image Processing

Background:

• Content identification systems rely on hash-based methods using perceptual hashing functions.
• Existing methods require manual adaptation of hashing functions to specific image manipulations, which is costly and time-consuming when new manipulations arise.

Purpose of the Study:

• To develop a novel approach for content identification systems that automatically learns an invariant feature space for hashing functions.
• To enhance the robustness and generalization capacity of image hashing against various manipulations.

Main Methods:

• Utilized self-supervised learning with a metric learning-based pretext task on unlabeled data.
• Enforced feature vector invariance against a set of image manipulations during training.
• Employed random sampling on the training set to expose the model to diverse perceptual information.

Main Results:

• Achieved excellent robustness against a comprehensive set of image manipulations, including horizontal flip and rotation.
• Demonstrated excellent identification performance and high discriminative ability against near-duplicate images.
• Showcased excellent generalization capacity, requiring no re-training or fine-tuning on new datasets.

Conclusions:

• The proposed self-supervised learning method effectively enables hashing functions to automatically learn invariant feature spaces for robust content identification.
• This approach significantly improves robustness, identification accuracy, and discriminative power, offering a highly generalizable solution.