Abstract

This paper proposes a realistic image splicing dataset named AbhAS for evaluating various image forensic algorithms. We evaluate the performance of our proposed AbhAS dataset against existing benchmark datasets by extracting high-energy coefficients from images belonging to each dataset with the application of Kekre and discrete cosine transforms (DCT). Thus, we obtain feature sets of sizes 12, 24, and 48 respectively which are passed through various machine learning classifiers. RandomForest (with DCT) and Bagging (with Kekre transform) provide the highest detection accuracy. We believe this dataset could add value to the existing work in the area of image forensics.

摘要

本文提出了一个名为 AbhAS 的逼真图像拼接数据集，用于评估各种图像取证算法。我们通过应用 Kekre 和离散余弦变换 (DCT) 从属于每个数据集的图像中提取高能系数来评​​估我们提出的 AbhAS 数据集与现有基准数据集的性能。因此，我们分别获得了大小为 12、24 和 48 的特征集，这些特征集通过了各种机器学习分类器。RandomForest（使用 DCT）和 Bagging（使用 Kekre 变换）提供最高的检测精度。我们相信这个数据集可以为图像取证领域的现有工作增加价值。