Image content identity verification with CNN considering cropping and compositing

抄録

Discriminative Perceptual Hashing (DPH) is a method that robustly verifies the identity of image content by fine-tuning a Convolutional Neural Network (CNN) for image classification to distinguish between target and non-target images. However, it does not explicitly address complex editing operations. This paper proposes a new DPH method that enables the model to verify the identity of target objects even after compositing-based modifications by adding images subjected to complex editing such as cropping the main object and pasting it into different contexts to the fine-tuning dataset. Furthermore, adjusting the types of editing operations used during training allows users to control the range of perceptual equivalence for copyright management. Experimental results demonstrate that the proposed method maintains the functionality of conventional methods and improves robustness against image compositing operations. This enhances the reliability of copyright verification under complex editing scenarios.