A novel reconstruction model for multi-frame super-resolution image based on lmix prior

A Robust Multiframe Image Super-Resolution Method in Variational Bayesian Framework

“…e initial values of the iteration process are chosen as follows: the initial motion vectors are estimated using the LK method [44], and assuming (Λ p k ) − 1 equal to zero, a noninformative hyperprior model assumption, i.e., a θ � b θ � 0, θ ∈ Θ, the HR image is initialized using bicubic interpolation of the LR reference frame; with the above initialization and Σ s k � 0 and Σ x � 0, the parameters w and v k and the distribution of β k are initialized using equation (28), and equations ( 23) and (24), respectively. en, α is initialized using equation (26), equation (29), and equation (33) for SAR, TV, and NS prior models, respectively. We used ‖x n − x n− 1 ‖ 2 2 /‖x n−1 ‖ 2 2 < 10 − 7 as convergence criterion, where x n and x n− 1 are acquired by using MAP to q n (x) and q n− 1 (x), respectively.…”

“…e minimization process of equation ( 20) has been described in [33], and it is summarized as below, for the n th iteration:…”

“…It is worth emphasizing here the profound influence made by our proposed observation model to the proposed SR algorithm. For the variational Bayesian SR algorithm in [33,34], only one parameter β k is used to quantify the noise level of k th LR image. Different from those algorithms, besides the parameter β k , local noise property of the LR images, indicated by the auxiliary vector v � v k 􏼈 􏼉 L k�1 in equation (23), is naturally introduced into the estimation process of q n (x) and q n (s k ).…”

“…Before comparing with other SR methods, we study the effect of the image prior model for our proposed observation model. Here, we compared three typical prior models, including SAR using equations ( 25) and ( 26), TV using equations ( 27)-( 29), and NS using equations ( 30)- (33), and denoted by La + SAR, La + TV, and La + NS, respectively. We evaluate the SR reconstruction on EIA and Lena for different noise models in cases where motion vector is available.…”

“…e image prior models have been extensively studied in the variational Bayesian framework [32][33][34]. However, all the mentioned works use a l2 normbased observation model.…”