物理模型的图像去雾算法MATLAB实现

物理模型的图像去雾算法MATLAB实现 结合了大气散射模型、暗通道先验和优化算法

---

一、算法原理与流程

1. 大气散射物理模型

\(I(x)=J(x)t(x)+A(1−t(x))\)

• \(I(x)\)：有雾图像
• \(J(x)\)：无雾图像（目标）
• \(t(x)\)：透射率
• \(A\)：大气光强度

2. 核心步骤

1. 暗通道估计：计算图像暗通道作为先验信息
2. 大气光估计：通过暗通道最亮区域确定全局大气光
3. 透射率优化：结合软抠图算法细化透射率图
4. 图像恢复：基于物理模型反演无雾图像

---

二、实现

function dehazed_img = physical_dehaze(I, block_size, w, t0)
    % 输入参数：
    % I: 输入有雾图像 (RGB)
    % block_size: 暗通道计算窗口尺寸
    % w: 透射率保留系数 (0<w<1)
    % t0: 透射率下限阈值

    % 预处理
    I = im2double(I);
    [rows, cols, ~] = size(I);
    
    % 1. 暗通道计算
    dark_channel = compute_dark_channel(I, block_size);
    
    % 2. 大气光估计
    A = estimate_atmospheric_light(I, dark_channel, block_size);
    
    % 3. 初始透射率估计
    transmission = 1 - w * dark_channel / A;
    transmission = max(transmission, t0);
    
    % 4. 透射率优化（软抠图）
    transmission = guided_filter(I(:,:,1), transmission, 60, 1e-3);
    
    % 5. 图像恢复
    J = zeros(size(I));
    for c = 1:3
        J(:,:,c) = (I(:,:,c) - A) ./ transmission + A;
    end
    
    % 后处理
    dehazed_img = post_process(J);
end

%% 暗通道计算函数
function dark = compute_dark_channel(I, block_size)
    [rows, cols, ~] = size(I);
    dark = zeros(rows, cols);
    
    % 最小值滤波
    half_block = floor(block_size/2);
    padded = padarray(I, [half_block, half_block], 'symmetric');
    
    for i = 1:block_size
        for j = 1:block_size
            window = padded(i:i+rows-1, j:j+cols-1, :);
            dark = min(dark, min(window, [], 3));
        end
    end
end

%% 大气光估计函数
function A = estimate_atmospheric_light(I, dark, block_size)
    [rows, cols, ~] = size(I);
    num_pixels = block_size^2;
    
    % 选择暗通道最亮的0.1%像素
    flat_dark = dark(:);
    [~, idx] = sort(flat_dark, 'descend');
    top_pixels = idx(1:round(0.001*num_pixels));
    
    % 在原图中取对应位置的最大值
    A = zeros(1,3);
    for c = 1:3
        channel = I(:,:,c);
        A(c) = mean(channel(top_pixels));
    end
end

%% 后处理函数
function J = post_process(J)
    % 直方图均衡化增强
    J = imadjust(J);
    
    % 自适应对比度拉伸
    J = adapthisteq(J, 'ClipLimit', 0.02);
    
    % 去雾增强
    J = imsharpen(J, 'Radius', 2, 'Amount', 1.5);
end

---

三、改进与优化

1. 自适应大气光估计

• 分块处理：将图像划分为多个区域独立估计大气光
• 惩罚因子：对过曝区域进行抑制

% 改进后的大气光估计
function A = adaptive_atmosphere(I, dark)
    [rows, cols, ~] = size(I);
    num_regions = 16;
    region_size = floor(rows/num_regions);
    
    A = zeros(1,3);
    for r = 1:num_regions
        for c = 1:num_regions
            % 区域定位
            y = (r-1)*region_size + 1;
            x = (c-1)*region_size + 1;
            
            % 区域暗通道统计
            region_dark = dark(y:y+region_size-1, x:x+region_size-1);
            mean_dark = mean(region_dark(:));
            
            % 选择最亮区域
            if mean_dark > 0.8
                A = A + mean(I(y:y+region_size-1, x:x+region_size-1), 'all');
            end
        end
    end
    A = A / num_regions;
end

2. 透射率优化策略

• 引导滤波：保持边缘信息的同时平滑噪声
• 多尺度处理：结合小波变换进行多分辨率优化

% 引导滤波优化透射率
function transmission = guided_filter_guide(I, transmission, radius, eps)
    mean_I = imgaussfilt(I, radius);
    mean_T = imgaussfilt(transmission, radius);
    
    corr_I = imgaussfilt(I.^2, radius) - mean_I.^2;
    corr_IT = imgaussfilt(I.*transmission, radius) - mean_I.*mean_T;
    
    var_I = corr_I + eps;
    cov_IT = corr_IT + eps;
    
    a = cov_IT ./ var_I;
    b = mean_T - a .* mean_I;
    
    transmission = a .* I + b;
    transmission = medfilt2(transmission, [3 3]);
end

参考代码 基于物理模型的图像去雾算法 www.youwenfan.com/contentcnl/79152.html

四、应用场景示例

1. 交通监控视频去雾

videoReader = VideoReader('highway.mp4');
videoWriter = VideoWriter('dehazed.avi');
open(videoWriter);

while hasFrame(videoReader)
    frame = readFrame(videoReader);
    dehazed = physical_dehaze(frame, 15, 0.92, 0.1);
    writeVideo(videoWriter, dehazed);
end
close(videoWriter);

2. 无人机航拍图像增强

% 加载航拍图像
img = imread('aerial_haze.jpg');

% 多尺度处理
pyramid = imagePyramid(img, 'NumLevels', 3);
dehazed_pyramid = cell(size(pyramid));
for i = 1:length(pyramid)
    dehazed_pyramid{i} = physical_dehaze(pyramid{i}, 9, 0.85, 0.15);
end

% 重建图像
dehazed_img = reconstructImage(dehazed_pyramid);

方法在标准测试集上表现优异，建议根据具体场景调整参数（如窗口尺寸、保留系数等），通过交叉验证优化模型性能。