Abstract
This paper presents a robust forensic approach for distinguishing between real-scene images and recaptured images—photographs of digital displays. The method leverages a combination of low-dimensional color moments in the HSV color space and statistical features derived from Discrete Cosine Transform (DCT) coefficients in both luminance and chrominance channels. By capturing the subtle physical distortions and double JPEG compression artifacts inherent in the recapture process, the proposed system achieves detection accuracies exceeding 96% across multiple datasets, significantly outperforming traditional texture-based forensic methods.
Introduction
The ubiquity of high-resolution digital cameras and displays has introduced a significant security vulnerability: image recapture. Malicious actors can photograph high-quality prints or LCD screens to deceive biometric authentication systems, such as facial recognition, or to generate fraudulent digital evidence. Unlike traditional image manipulations like splicing or cloning, a recaptured image is a legitimate photograph taken by a physical camera. Consequently, it often evades standard forensic tests designed to detect software-based editing.
However, the physical process of recapture introduces unique artifacts. The ambient lighting of the environment interacts with the light emitted by the source medium, and the image typically undergoes a second round of JPEG compression. This research aims to identify and extract these specific "fingerprints" to reliably differentiate a direct capture of a real-world scene from a capture of a digital screen.
Methodology
The authors propose a dual-feature extraction framework, utilizing a Support Vector Machine (SVM) for final classification.
Color Moment Features
To capture shifts in color diversity caused by the recapture environment and the limited gamut of digital screens, the image is first converted from the RGB to the HSV (Hue, Saturation, Value) color space. The first three central moments—mean, standard deviation, and the third root of skewness—are calculated for each of the three channels. This results in a compact 9-dimensional feature vector. These moments effectively characterize the "color anomalies" introduced during recapture without the computational overhead of high-dimensional color histograms.
Mode-Based First Digit Features (MBFDF)
Recaptured images generally undergo double JPEG compression, leaving statistical traces in their DCT coefficients. The authors extract the probability distribution of the first digits of quantized DCT coefficients from both the luminance (Y) and chrominance (Cr) components. By analyzing the first five "modes" (Zigzag scan positions) across both channels, a 90-dimensional feature vector is generated. A key contribution of this work is the inclusion of the chrominance component, which is often overlooked in compression-based forensics but proves highly effective here.
Experimental Results
The algorithm was evaluated on three distinct datasets ranging from small-resolution sets to high-quality captures from Canon and Nikon cameras. The combined feature set (Color Moments + MBFDF) consistently outperformed texture-based methods like Local Binary Patterns (LBP). Accuracy reached 93.36% on Dataset 1 and peaked at 96.70% on Dataset 3. Notably, the system demonstrated high sensitivity, with recall rates often exceeding 97%, indicating a low rate of false negatives for recaptured images.
Discussion and Conclusion
The study highlights that the chrominance (Cr) component in the DCT domain is particularly sensitive to the recapture process. Furthermore, the analysis of color moments revealed that recaptured images tend to have higher mean values but lower variance and skewness compared to real-scene images, likely due to the dynamic range limitations of LCD screens.
While highly effective for LCD-based recapture, future work should address performance on high-quality paper prints and under extreme environmental lighting conditions. Additionally, investigating anti-forensic techniques that might attempt to mask these color and compression signatures remains a critical area for further research.
Summary updated