Algorithm 1.

Our learning method for active patch dictionary

Initialization:
1:	Get 𝒟o from clustering algorithm;
2:	𝒟f ← ∅; ∀Mbest(I, pos) ← 0; ∀pr ← void, pr ∈ ∀IR(I);
Process:
3:	Sort 𝔸(k) ∈ 𝒟o w.r.t. F(𝔸(k)) in descending order;
4:	for each 𝔸(k) ∈ 𝒟o do
5:	∀I ∈ Ω, record current IR(I) and Mbest(I, pos); ΔE ← 0;
6:	for each position pos of each image I ∈ Ω where 𝔸(k) is firing do
7:	for each pixel $p_d\in S_{d_t}^{\mathit{\text{best}}},S_{d_t}^{\mathit{\text{best}}}\in {𝕊}_D^{(k)}$ do
8:	Calculate ΔE according to Equ. 4;
9:	if the corresponding pr = void or M(𝔸(k), I, pos) > Mbest(I, pos) then
10:	pr ← pd; Mbest(I, pos) ← M(𝔸(k), I, pos);
11:	end if
12:	end for
13:	end for
14:	if ΔE < 0 then
15:	Add 𝔸(k) to 𝒟f
16:	else
17:	∀I ∈ Ω, restore IR(I) and Mbest(I, pos) to previous value;
18:	end if
19:	end for