Phi-divergence

测度的绝对连续性(Absolute continuity of measures)

定义. 假设 B \mathcal{B} B 是定义于 X X X 的子集上的一个 σ \sigma σ-代数, μ , ν \mu, \nu μ,ν B \mathcal{B} B 上的两个测度, 如果对于任意满足 μ ( A ) = 0 \mu(A)=0 μ(A)=0 的子集 A ∈ B A\in\mathcal{B} AB,有 ν ( A ) = 0 \nu(A)=0 ν(A)=0,则我们称 ν \nu ν 相对于 μ \mu μ绝对连续的。

Radon-Nikodym定理. 如果测度 ν \nu ν 相对于 μ \mu μ 是绝对连续的,那么存在一个函数 f ∈ L 1 ( μ ) f\in L^1(\mu) fL1(μ) 使得 ν = f μ \nu=f\mu ν=fμ,i.e.,
ν ( A ) = ∫ A f d μ ∀ A ∈ B \nu(A)=\int_A fd\mu \qquad \forall A\in\mathcal{B} ν(A)=AfdμAB

ϕ \phi ϕ-divergence

在这里插入图片描述

请学习我给出的代码,继续优化并给出修改后完整的matlab代码:clc; clear; close all; % 参数设置 M = 1024; N = 1024; file = '1.mat'; % 输入你的全息图 data = load(file); fn = fieldnames(data); holo= double(data.(fn{1})); holo_amp = sqrt(holo / max(holo(:))); % 归一化幅值 amp_gt = im2double(imread('BJUT.jpg'));% 原始复振幅的“幅值图” amp_gt = imresize(amp_gt, [1024, 1024]); amp_gt = rgb2gray(amp_gt); amp_gt = amp_gt / max(amp_gt(:));%归一化幅值 lambda = 532e-9; % 波长 dx = 2.9e-6; % 像素尺寸 z = 27e-3; % 传播距离 N_iter = 100; % 迭代次数 % 相位复原 [U_obj, recon_amp, recon_phase] = phase_recovery(holo_amp, lambda, dx, z, N_iter, amp_gt); % 显示结果 figure; subplot(1,2,1); imshow(recon_amp, []); title('复原幅值'); subplot(1,2,2); imshow(recon_phase, []); title('复原相位'); % 评估指标 recon_amp_norm = recon_amp / max(recon_amp(:)); mse = mean((amp_gt(:) - recon_amp_norm(:)).^2); rmse_val = sqrt(mse); psnr_val = 10 * log10(1 / mse); fprintf('[评估结果] PSNR = %.2f dB,RMSE = %.4f\n', psnr_val, rmse_val); %% === 角谱传播函数 === function U_out = angular_spectrum(U_in, z, wavelength, dx) [M, N] = size(U_in); k = 2 * pi / wavelength; fx = (-N/2:N/2-1) / (N * dx); fy = (-M/2:M/2-1) / (M * dx); [FX, FY] = meshgrid(fx, fy); H = exp(1j * k * z * sqrt(1 - (wavelength * FX).^2 - (wavelength * FY).^2)); H((wavelength * FX).^2 + (wavelength * FY).^2 > 1) = 0; U_out = ifftshift(ifft2(fft2(fftshift(U_in)) .* fftshift(H))); end %% === 相位复原主函数 === function [U_obj, recon_amp, recon_phase] = phase_recovery(holo_amp, wavelength, dx, z, num_iter, amp_gt) [M, N] = size(holo_amp); H = holo_amp; % 初始记录面幅值 phase_est = zeros(M, N); % 初始相位设为 0 U_rec = H .* exp(1j * phase_est); % 初始复振幅 for iter = 1:num_iter U_obj = angular_spectrum(U_rec, -z, wavelength, dx); % 传播到物面 A = abs(U_obj); phi = angle(U_obj); A = min(A, 1); % 正吸收约束 U_obj = A .* exp(1j * phi); % TV 去噪(Split Bregman) U_obj = TV_denoise(U_obj, 0.02, 5); % 可调参数 % 传播回记录面并施加幅值约束 U_rec = angular_spectrum(U_obj, z, wavelength, dx); U_rec = H .* exp(1j * angle(U_rec)); % 可视化中间迭代 if mod(iter, 20) == 0 figure(2); imshow(abs(U_obj), []); title(['Iter ' num2str(iter)]); drawnow; end end % 最后再传播一次获得最终复原图 U_obj = angular_spectrum(U_rec, -z, wavelength, dx); recon_amp = abs(U_obj); recon_phase = angle(U_obj); end %% === Split Bregman TV 去噪 % === function U_out = TV_denoise(U_in, lambda, iter_tv)%lambda:TV正则项系数 [m, n] = size(U_in);%初始化 U_out = U_in; dx = zeros(m, n); dy = zeros(m, n); bx = zeros(m, n); by = zeros(m, n); mu = 1; for k = 1:iter_tv [Ux, Uy] = gradient(U_out);%计算梯度项 dx = shrink(Ux + bx, lambda / mu);%更新 d(软阈值) dy = shrink(Uy + by, lambda / mu); div_d_b = divergence(dx - bx, dy - by);%更新 U(解一个波动方程) U_out = solve_poisson(U_in + mu * div_d_b, mu); [Ux, Uy] = gradient(U_out);%更新 Bregman 参数 bx = bx + (Ux - dx); by = by + (Uy - dy); end end function out = shrink(x, t)%软阈值函数,适用于复数 out = max(abs(x) - t, 0) ./ (abs(x) + 1e-8) .* x; end function div = divergence(px, py)%计算散度 [m, n] = size(px); div = zeros(m, n); div(:,1:end-1) = px(:,1:end-1) - px(:,2:end); div(:,end) = px(:,end); div(1:end-1,:) = div(1:end-1,:) + py(1:end-1,:) - py(2:end,:); div(end,:) = div(end,:) + py(end,:); end function U = solve_poisson(rhs, mu)%使用频域方法解泊松方程 [m, n] = size(rhs); [fx, fy] = meshgrid(0:n-1, 0:m-1); fx = fx / n; fy = fy / m; denom = 1 + mu * ((2 - 2 * cos(2 * pi * fx)) + (2 - 2 * cos(2 * pi * fy))); U = real(ifft2(fft2(rhs) ./ denom)); end
最新发布
03-21
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包

打赏作者

zte10096334

你的鼓励将是我创作的最大动力

¥1 ¥2 ¥4 ¥6 ¥10 ¥20
扫码支付:¥1
获取中
扫码支付

您的余额不足,请更换扫码支付或充值

打赏作者

实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值