Source camera identification model: Classifier learning, role of learning curves and their interpretation

摘要

Source camera identification is the problem of associating an image with its source device. Majority of the existing source detection techniques have their operations based on machine learning principles, and report a considerably high accuracy as far as prediction is concerned. Such techniques follow a basic operating principle: extract appropriate features from images, train classifier for camera prediction, predict the image source class. In the source camera identification problem, the tolerance for false acceptance rate is extremely low, ideally zero. Hence, it is imperative that the model built should predict the source of unknown data with high accuracy. In this scenario, the learning process that a model undergoes, plays the most crucial role, and subsequently affects the accuracy of prediction majorly. In this paper, we discuss various techniques to make an image source identification model learn properly, and establish the importance of concentrating on learning part of a system, through proper interpretation of learning curves. We tested the approaches on the Dresden image database. Our experimental results prove that in this field of research, for fair evaluation and comparison of state-of-the-art techniques, the use of credible benchmark database as Dresden is uncompromisable, as compared to proprietary image datasets.