基于样本错配训练的图像PRNU噪声提纯方法

Purification Scheme of Image PRNU Noise Based on Cross⁃Matching Sample Training

基于传感器模式噪声的图像来源鉴别算法的关键在于如何获取高质量的光响应非均匀性(Photo-response non-uniformity,PRNU)噪声,目前大多数增强PRNU噪声质量的算法以及出于实际应用的目的对其压缩的算法几乎是在人工假设模型的基础上实现的。本文提出了一种基于样本错配训练的图像PRNU噪声提纯方法,该方法使用了深度堆叠自编码器网络,并设计了一种样本错配的技术对其进行训练。这种基于样本错配训练的端到端深度神经网络的使用有效避免了人工设计算法的局限性,对图像PRNU噪声进行了有效提纯,进而提升了其用于图像来源鉴别时的性能。在Dresden图像数据集上进行的比较实验结果表明,提纯后的PRNU噪声具有更好的性能。

The key to the image source identification algorithm based on sensor pattern noise is how to obtain high-quality PRNU noise.At present,most of the algorithms include to improve the quality of PRNU noise and compress them for purposes of practical application based on artificially assumed models.This paper proposes a purification scheme of image PRNU noise based on cross-matching sample training.The scheme uses a deep stacked auto-encoder network and designs a technology of sample cross-matching to train it.The use of this end-to-end deep neural network based on cross-matching sample training effectively avoids the limitations of manual design algorithms and purifies the image PRNU noise.Then the performance of PRNU noise is improved for image source identification.The results of the comparative experiment on the Dresden image dataset show that the purified PRNU noise has better performance.