CNN-based perceptual hashing scheme for image groups suitable for security systems
Sugawara Yusei, Zhaoxiong Meng, Morizumi Tetsuya, Sumiko Miyata, Hosono Kaito, Hirotsugu Kinoshita · 2023
Perceptual hashing, which generates a message digest showing how humans perceive similarity in images, is suitable for ensuring the equivalence between a modified/edited image and the original. Conventional perceptual hashing is mainly utilized for similarity-based image retrieval and is not appropriate for image identification, which is required in security systems such as digital rights management. We previously developed a construction method for perceptual hashing in security systems that utilizes a convolutional neural network (CNN). In practical applications, multiple different images are published in various media forms (e.g., articles or books), so generating an identical message digest for each of these images simultaneously makes content management easier. Therefore, in this work we extend our earlier CNN-based perceptual hashing scheme so that it can generate an identical message digest for images in a group. This approach reduces the computational cost for fine-tuning CNN compared to generating a perceptual hash for each image in a group individually.