This metric evaluates the stability of fine details by monitoring the evolution of edge gradients. Successful numeric scores are official: they enter the quality dimension and the public leaderboard vector. We propose a vectorized Tenengrad method: independently compute the sum of absolute gradients in the horizontal (Gx) and vertical (Gy) directions, and construct a 2-D sharpness vector to capture texture-direction features:
gt=(∑∣Gx∣∑∣Gy∣)(6)
where Gx and Gy are Sobel responses (3×3 kernel) on the grayscale frame It.
A noise-aware circuit breaker distinguishes true detail from high-frequency artifacts. Let nt be the BRISQUE noise score at frame t and τ a noise threshold. When high noise (nt>τ) occurs for 5 consecutive frames, the triggered state is Itrig=1. The fusing penalty logic M combines texture similarity with the breaker state (similarity S(⋅,⋅) follows the cosine inner product in Eq. (3)):
BRISQUE is undefined for a zero-variance frame. Such a frame receives the configured worst-case brisque_bad value and remains in the aggregate, so a collapse to a flat or black frame is penalized rather than silently excluded as a missing score.
An upper-convex logarithmic transformL(⋅) enhances low-to-mid score segments (x∈[0,1], avoiding undefined logarithms and preserving normalization):
L(x)=ln(1+k)ln(1+k⋅x),k>0(10)
where k controls the upper-convex degree; k∈[10,20] is common in vision pipelines. The numerator ensures x=0⇒L(x)=0; the denominator ensures x=1⇒L(x)=1.
Temporal aggregation uses normalized exponentially decaying weights w^(d) (Eq. (5) in the brightness-consistency metric, d∈[1,T−1]). The comprehensive score is:
SSharpness=d=1∑T−1w^(d)⋅L(M(xd+1))(12)
where S(gt,g1) is the cosine similarity between the sharpness vector of frame t and the initial frame. This mechanism distinguishes genuine detail from high-frequency noise and penalizes systematic visual collapse.
