结合深度迭代缩放卷积神经网络的PRNU提取算法

光响应非均匀性（photo-response non-uniformity,PRNU）是用于数字图像设备溯源的一种重要特征,也被称为成像设备指纹。针对图像真实噪声包含PRNU和大量未知噪声的复杂特性,本文提出一种结合深度迭代缩放卷积神经网络的PRNU数字成像设备指纹提取算法。首先,通过连续重复的缩小与放大特征图的分辨率来提高GPU内存利用效率和生成大的感受野,尽可能的提取包含完整PRNU指纹的真实噪声。然后,利用来自同一数字成像设备多幅图像的噪声残差来估计PRNU指纹。本文算法在相机溯源数据集Dresden和手机溯源数据集Daxing上进行了测试。实验结果显示,与传统方法相比本文算法具有的更好的识别率和普适性。 

PRNU Extraction Algorithm Combined with Deep Iterative Down-Up CNN Model

Photo-response non-uniformity (PRNU) referred as device fingerprint can be used for digital imaging device source identification. In view of the complexity of PRNU and a large number of unknown noises in the real image noise, this paper proposed a PRNU extraction algorithm combined with deep iterative down-up convolutional neural network (CNN) model. In particular, in order to extract the complete PSNR fingerprint from the real image noise, the resolution of the feature map was decreased and increased repetitively to efficiently employ graphics processing unit (GPU) memory and yield large receptive fields. Then, the noise residual was computed by multiple images from the same digital imaging device, which was utilized to estimate the PRNU fingerprint. The proposed algorithm was evaluated by Dresden camera database and Daxing smartphone database, and the experimental results show that the proposed algorithm achieves better recognition accuracy and universality compared with traditional methods.