CFA插值中预测误差方差在图像取证中的应用

Forecast Error Variance Analysis of CFA Interpolation and its Application in the Image Forensics

针对数字图像取证中自然图像(PIM)和计算机生成图像(PRCG)识别方案的特征维数高、通用性差等问题,提出一种基于滤色器阵列(CFA)插值中预测误差方差分析的图像取证方案.首先,对CFA插值过程中的预测误差方差进行傅里叶谱分析,根据是否存在明显的周期性峰值现象来区分PIM和PRCG;然后,对傅里叶谱中周期性峰值模型进行分析,根据峰值特征来识别PIM的来源设备;最后,在哥伦比亚大学自然图像和计算机生成图像数据库ADVENT上进行实验,结果表明,该方案能够精确区分PIM和PRCG,对PIM来源设备(佳能、尼康和索尼)的识别率可高达93%.

For the issues that the photographic image（PIM） and computer generated images（PRCG） identification scheme have features of poor generality and high dimension in image forensics, an image forensics scheme base on forecast error variance analysis in color filter array（CFA） interpolation is proposed. First, the Fourier spectrum of prediction error variance of CFA interpolation is analyzed, and the PIM and PRCG are distinguished according to whether there is a distinct periodic peak phenomenon. Then, the periodic peak model is analyzed, and the source of PIM is identified according to the peak value features. Finally, experiments have been done on natural images from Columbia University and computer generated image database ADVENT. Experimental results show that the proposed scheme can accurately distinguish between PIM and PRCG, and the recognition rate of the PIM source devices（Canon, Nikon and SONY） reached 93%.