Design scheme of perceptual hashing based on output of CNN for digital watermarking

Perceptual hashing generates a message digest based on the image content of the human visual system, and differs from the cryptographic hash function that generates the hash value based on each bit of the image file. We apply perceptual hashing to digital watermarking to generate watermark information after each image modification/editing, and verify that modified/edited images and the original image are the same in copyright. To obtain a stable perceptual hash value robust to image modification/editing for digital watermarking, we previously developed a construction method for perceptual hashing using a convolutional neural network (CNN). This was necessary because the conventional perceptual hash algorithms are used for database retrieval, and the required characteristics are different from those used for digital watermarking. However, in this method we needed to fine-tune the CNN for each image used to calculate the perceptual hash value, which led to inefficiency. In order to make the calculation of the perceptual hash value more efficient, we propose a construction method for perceptual hashing based on CNN that does not require fine-tuning. In the proposed method, an image is input to the CNN and the perceptual hash value is calculated based on the response of the output layer of the trained CNN.