Enhancing PRNU-based camera identification by denoising filters and similarity metrics

Source camera identification through photo response nonuniformity (PRNU) analysis is a well-established forensic technique, yet the choice of experimental conditions and parameters of the algorithms critically impacts identification reliability. For the first time, this study systematically evaluates 11 denoising filters (linear, nonlinear, and adaptive) for noise pattern extraction in conjunction with six similarity image metrics to establish an optimal framework for PRNU extraction and matching. To measure the PRNU patterns and analyze the identification reliability, a dataset of 4387 light uniform, dark, and natural images of three cameras was created (LDN dataset). Among similarity metrics, normalized standard deviation (NSTD) substantially outperforms Pearson's correlation coefficient (7.6× higher value of the comparison parameter), while peak-to-correlation energy, phase-only correlation, and structural similarity index measure exhibit negligible comparative utility. The Lee filter-originally designed for multiplicative speckle noise reduction-achieves an 8.6× higher comparison parameter over the Wiener filter and 12.1× higher comparison parameter over the median filter. This research establishes a comprehensive experimental framework for increasing the robustness of device-level identification in forensic investigations.