Feature Extraction Suitable for Double JPEG Compression Analysis Based on Statistical Bias Observation of DCT Coefficients

Photographs taken by smartphones and camera devices are generally compressed using JPEG by default when they are saved. If such an image is edited, it is decompressed and processed, and then recompressed through JPEG. Therefore, an edited image must be compressed by JPEG more than once. Using this characteristic, a forensic technique has been studied to detect image tampering by detecting distortions caused by double compression. In our previous study, to analyze the JPEG compression history using a convolutional neural network and (CNN), we observed a histogram calculated from the low-frequency components in
$8\times 8$
sized blocks of images having a pixel resolution of
$512\times 512$
. However, there have been no detailed considerations regarding the range of observed histograms or the selection of DCT coefficients used to extract the features from a given image. In this study, we first examine the range of his-tograms to measure the usefulness of the classification of double JPEG-compressed images, and then examine the classification accuracy by increasing the number of DCT coefficients observed in the low-to mid-frequency components. Our experiment results indicate that [−40,40] is an appropriate range for observing a histogram, and the selection of DCT coefficients strongly depends on the image size because of the difference in the amount of useful statistical information available.