Normalization Video Fidelity Metric (NVFM)

NVFM is implemented based on a visibility prediction employing a normalization mode. The perceptual decomposition is done separately in time and space in the pixel domain, without using Fourier transforms. Spatial (or space) decomposition uses the steerable pyramid [135]. The transform decomposes an image into spatial frequency and orientation bands. It is implemented as a recursive pyramidal scheme. The number of bands has been chosen to be equal to the Gabor implementation, as in MPQM. The orientation decomposition is steerable for each frequency band to be invariant by rotation. A temporal filter bank is used to do the temporal decomposition. The bank also has to approximate the Gabor temporal bank using infinite impulse response (IIR) filters. Then, an excitatory-inhibitory stage is used to measure the energy by squaring the output of the linear transform stage. After that, a normalization stage consisting of calculating the ratio between the excitatory and inhibitory mechanisms is carried out. The final measure of the metric is a squared-error norm, calculated as the squared vector sum of the difference of the normalized responses. Let $R_o$ be a vector collecting the sensor outputs for the original sequence and $R_d$ be that of the distorted one. The detection mechanism computes the vector distance $\triangle R$ using

$$\triangle R=\vert R_o-R_d\vert^2.$$