High Dynamic Range Video

1.8 Compositing as Bayesian Joint Estimation

Up to this point in the discussion, we have not characterized the noise terms inherent in the imaging process, nor have we used them to improve the estimation of the light quantity present in the original scene. In this section we develop a technique to incorporate a noise model into the estimation process.

Our approach for creating an HDR image from N input LDR images begins with the construction of a notional N-dimensional inverse camera response function that incorporates the different exposure and weighting values between the input images. Then we could use this to estimate the photoquantity $\widehat{q}$ at each point by writing $\widehat{q}(\mathbf{x})=f^{-1}(f_1,f_2,\dots ,f_N)/k_1$. In this case f⁻¹ is a joint estimator that could ...