主权项 |
一种基于压缩感知的影像两步重构法,其特征是:所述方法的过程为:步骤一:输入测量矩阵Φ,稀疏变换基Ψ,时相t1和t2的测量数据Y<sub>t1</sub>、Y<sub>t2</sub>,时相t1的影像X<sub>t1</sub>,其中Φ∈R<sup>M×N</sup>,Ψ∈R<sup>N×N</sup>,Y<sub>t1</sub>,Y<sub>t2</sub>,ΔY∈R<sup>N×L</sup>,X<sub>t1</sub>,X<sub>t2</sub>,ΔX∈R<sup>N×L</sup>,C<sub>t1</sub>,C<sub>t2</sub>,ΔC∈R<sup>N×L</sup>,M<N,j∈[1,L],令Y<sub>t1</sub>=ΦX<sub>t1</sub>,Y<sub>t2</sub>=ΦX<sub>t2</sub>,ΔY=Y<sub>t2</sub>‑Y<sub>t1</sub>,ΔX=X<sub>t2</sub>‑X<sub>t1</sub>,ΔC=C<sub>t2</sub>‑C<sub>t1</sub>,ΔY=ΦΔX,ΔY=ΦΨ′ΔC;步骤二:采用OMP(或其他l<sub>0</sub>算法)基于min||Δx<sub>j</sub>||<sub>0</sub> s.t. Δy<sub>j</sub>=ΦΔx<sub>j</sub>逐列重构ΔX的各列,取得每列的重构结果Δx<sub>Rj</sub>和ΔX的重构结果ΔX<sub>R</sub>;步骤三:设定临界值cri1,判断ΔX<sub>R</sub>各列的<img file="FDA0000726149090000012.GIF" wi="529" he="118" />如果是,则属于不确定区域;否则属于确定区域。步骤四:提取ΔX中不确定区域列号的集合J,在第j次迭,如果j∈J,那么采用OMP(Orthogonal matching pursuit)(或其他l<sub>0</sub>算法,例如:ROMP(Regularized orthogonal matching pursuit)、StOMP(Stagewise orthogonal matching pursuit)、SP(Subspace pursuit)、CoSaMP(Compressive sampling matching pursuit)、MP(Matching pursuit))基于min||c<sub>j</sub>||<sub>0</sub> s.t. Δy<sub>j</sub>=ΦΨ′Δc<sub>j</sub>或采用BP(Basis Pursuit)(或其他l<sub>1</sub>算法,例如:GPSR(Gradient projection for sparse reconstruction))基于min||Δc<sub>j</sub>||<sub>1</sub>s.t.Δy<sub>j</sub>=ΦΨ′Δc<sub>j</sub>重构ΔX的该列Δx<sub>j</sub>的稀疏变换域系数Δc<sub>j</sub>,然后解算Δx<sub>Rj</sub>=Ψ′Δc<sub>Rj</sub>;或采用TV(minimization of total variation)算法基于min||Δx<sub>j</sub>||<sub>TV</sub> s.t.Δy<sub>j</sub>=ΦΔx<sub>j</sub>重构ΔX的该列Δx<sub>j</sub>,最后取得该列的重构结果Δx<sub>Rj</sub>;步骤五:输出重构信号ΔX<sub>R</sub>和时相t2的影像X<sub>t2</sub>=X<sub>t1</sub>+ΔX<sub>R</sub>。 |