Block-Based MAP Superresolution Using Feature-Driven Prior Model

Algorithm 1

The summary of AR-BMSFP Algorithm.

Objective: Estimate super-resolution 𝑥.

Input:

  Low-resolution images sequence {𝑦_𝑡}^𝑇_𝑡=1, Motion matrix 𝐹_𝑡, Border-expansion width 𝑤, Total rows of block 𝑝, Total

  columns of block 𝑞

Initialization: Estimate initial image 𝑥⁰ with shift and add Algorithm

Splitting:

  Split the initial image 𝑥⁰ into 𝑝×𝑞 blocks

  Expand each block 𝑤 pixels width and obtain 𝑥⁰_𝑖

  Estimate 𝜎_𝐿𝑖, 𝜎_𝐺𝑖 and 𝑅_𝑖 for each block 𝑥⁰_𝑖

  Compress 𝐹_𝑡 into 𝐹_𝑡𝑖 according to 𝑥⁰_𝑖

Optimization:

  Form the block cost function 𝐶_𝑖 using (18)

  Optimize 𝐶_𝑖 with Scaled Conjugate Gradients (SCG) and initial block 𝑥⁰_𝑖 obtaining the overlapped super-resolution

  block 𝑥^∞_𝑖

Recombination:

  Cut border of 𝑥^∞_𝑖 with 𝑤 pixels width, obtaining super-resolution block 𝑥_𝑖; Combine all super-resolution blocks obtaining

  the result image 𝑥

Result: The output is 𝑥.