Hypothesis Testing in Camera Attribution
Definition
The framework in which H0 is 'the query image was not taken by this camera' and H1 is 'it was'. The PCE or NCC value is compared to a threshold calibrated from empirical distributions of true and false matches, and a decision is made with stated false-positive and false-negative rates.
- H0
- The query image was not taken by this camera
- H1
- The query image was taken by this camera
- Statistic compared
- PCE or NCC value against a calibrated threshold
- Output
- Decision with stated false-positive and false-negative rates
Common questions
Why is a calibrated threshold necessary rather than just comparing PCE values directly?+
The distribution of PCE values for true matches and false matches overlaps somewhat, so the threshold is set from empirical data on known-match and known-non-match image pairs to control the trade-off between wrongly excluding the true camera and wrongly including an innocent one.
Why does the reported false-positive rate matter for court use of PRNU camera attribution?+
It quantifies how often the test would falsely link an image to a camera that did not take it under similar conditions, letting the court weigh the strength of a reported match rather than treating any PCE value above threshold as certain identification.
Related terms
- BM3D (Block-Matching 3D)
- A state-of-the-art denoising algorithm that groups similar image patches across the image, transforms them jointly, thresholds in the transform domain, and inverts....
- NCC (Normalised Cross-Correlation)
- The standard measure of similarity between a query noise residual and a reference PRNU pattern. It produces a value between -1 and...
- Noise Residual (W_hat)
- The per-image estimate of PRNU, computed by subtracting a denoised version of the image from the original. It contains PRNU plus random...
- PCE (Peak-to-Correlation Energy)
- A refined statistic that divides the peak of the cross-correlation surface by the energy of the non-peak region. PCE is more sensitive...
- Wavelet Denoiser
- A denoising filter that works in the wavelet transform domain, thresholding coefficients at each scale and direction to remove random noise while...