ISO Setting Estimation Based on Convolutional Neural Network and its Application in Image Forensics

摘要

The ISO setting, which is also known as film speed, influences the noise characteristics of output images. As a consequence, it plays an important role in noise based forensics. Whenever the ISO setting information cannot be retrieved from the image metadata, estimating the ISO setting of a probe image from its content is of forensic significance. In this work, we propose a convolutional neural network, called ISONet, for ISO setting estimation. The proposed ISONet can successfully infer the ISO setting both globally (image-level) and locally (patch-level). It not only works on uncompressed images, but also is effective on JPEG compressed images. We apply the ISONet on two typical forensic scenarios, one is the image splicing localization and the other is the Photo Response Non-Uniformity (PRNU) correlation prediction. A series of experiments show that the ISONet can yield a remarkable improvement in both forensic scenarios.