The Randlet Transform Applications to Universal Perceptual Hashing and Image Identification

We introduce a new transform called the randlet transform and explore applications to universal perceptual image hashing (key-based randomized digests) and image identification. Our transform yields robust signal representations that are nearly invariant to several perceptually insignificant transformations (attacks, intentional or otherwise). Our signal representation is hard to guess without the secret key used in its derivation and is motivated by applications such as image identification and watermarking where attack resistance is important. It is also interesting in itself as a transform and in regular signal processing tasks such as scene detection, motion compensation algorithms, compression, etc. We consider the performance of the randlet transform and the wavelet transform against attacks based on rotation, cropping, scaling, additive noise, and JPEG compression. The randlet transform achieves superior performance to the wavelet transform in the task of image identification. For example, a randlet transform with a false positive rate of 2% can detect a 10-degree rotation with a false negative rate of 3.2%, while the best wavelet transform with the same false positive rate has a false negative rate of 38.2%.