VQA_model/compute_correlation.m

%% 计算视频质量特征与MOS的相关性
% 此脚本用于计算已提取的视频质量特征与人工评分(MOS)之间的相关性

clear all;
clc;

% 加载提取的特征
fprintf('加载特征数据...\n');
load('video_quality_features.mat');

% 加载或读取MOS数据 - 这里假设scoreList已经包含在特征文件中
% 如果需要从另一个文件加载，取消下面的注释并调整路径
% mos_file = '/path/to/mos_file.txt';
% mos_data = readtable(mos_file);
% scoreList = mos_data.score;

% 检查各特征向量是否有效
fprintf('检查特征有效性...\n');
valid_indices = ~isnan(BRISQUEall) & ~isnan(scoreList);
if sum(valid_indices) < 10
    warning('有效样本太少，无法计算可靠的相关性！');
    return;
end

% 移除无效样本
BRISQUEall_valid = BRISQUEall(valid_indices);
BMPRIall_valid = BMPRIall(valid_indices);
BPRIall_valid = BPRIall(valid_indices);
HOSAall_valid = HOSAall(valid_indices);
NIQEall_valid = NIQEall(valid_indices);
QACall_valid = QACall(valid_indices);
scoreList_valid = scoreList(valid_indices);

% 初始化结果表格
methods = {'BRISQUE', 'BMPRI', 'BPRI', 'HOSA', 'NIQE', 'QAC'};
metrics = {'SRCC', 'KRCC', 'PLCC', 'RMSE'};
results = zeros(length(methods), length(metrics));

% 计算相关性
fprintf('计算相关性指标...\n');

% 计算BRISQUE相关性
results(1, :) = compute_metrics(BRISQUEall_valid, scoreList_valid);

% 计算BMPRI相关性
results(2, :) = compute_metrics(BMPRIall_valid, scoreList_valid);

% 计算BPRI相关性
results(3, :) = compute_metrics(BPRIall_valid, scoreList_valid);

% 计算HOSA相关性
results(4, :) = compute_metrics(HOSAall_valid, scoreList_valid);

% 计算NIQE相关性
results(5, :) = compute_metrics(NIQEall_valid, scoreList_valid);

% 计算QAC相关性
results(6, :) = compute_metrics(QACall_valid, scoreList_valid);

% 显示结果
fprintf('\n==== 相关性分析结果 ====\n');
result_table = array2table(results, 'RowNames', methods, 'VariableNames', metrics);
disp(result_table);

% 保存结果
save('correlation_results.mat', 'result_table', 'results', 'methods', 'metrics');
fprintf('结果已保存到 correlation_results.mat\n');

% 绘制散点图
figure('Position', [100, 100, 800, 600]);
best_method_idx = find(results(:, 1) == max(results(:, 1)));
best_method = methods{best_method_idx};
best_features = [];

switch best_method
    case 'BRISQUE'
        best_features = BRISQUEall_valid;
    case 'BMPRI'
        best_features = BMPRIall_valid;
    case 'BPRI'
        best_features = BPRIall_valid;
    case 'HOSA'
        best_features = HOSAall_valid;
    case 'NIQE'
        best_features = NIQEall_valid;
    case 'QAC'
        best_features = QACall_valid;
end

% 执行非线性拟合
beta_init = [max(scoreList_valid), min(scoreList_valid), mean(best_features), 0.5];
[beta, ~] = nlinfit(best_features, scoreList_valid, @logistic_func, beta_init);
fitted_mos = logistic_func(beta, best_features);

% 绘制散点图和拟合曲线
scatter(best_features, scoreList_valid, 50, 'filled');
hold on;
x_range = linspace(min(best_features), max(best_features), 100);
y_fit = logistic_func(beta, x_range);
plot(x_range, y_fit, 'r-', 'LineWidth', 2);
title(sprintf('%s 方法与MOS的散点图及拟合曲线', best_method), 'FontSize', 14);
xlabel('特征值', 'FontSize', 12);
ylabel('MOS值', 'FontSize', 12);
grid on;
legend('原始数据', '拟合曲线');
saveas(gcf, 'best_method_scatter.png');
fprintf('散点图已保存到 best_method_scatter.png\n');

%% 辅助函数
function metrics = compute_metrics(predicted, actual)
    % 计算SRCC
    srcc = corr(predicted, actual, 'Type', 'Spearman');
    
    % 计算KRCC
    krcc = corr(predicted, actual, 'Type', 'Kendall');
    
    % 执行非线性拟合
    beta_init = [max(actual), min(actual), mean(predicted), 0.5];
    try
        [beta, ~] = nlinfit(predicted, actual, @logistic_func, beta_init);
        fitted_mos = logistic_func(beta, predicted);
        
        % 计算PLCC
        plcc = corr(fitted_mos, actual, 'Type', 'Pearson');
        
        % 计算RMSE
        rmse = sqrt(mean((fitted_mos - actual).^2));
    catch
        warning('非线性拟合失败，使用线性相关系数代替');
        plcc = corr(predicted, actual, 'Type', 'Pearson');
        rmse = sqrt(mean((predicted - actual).^2));
    end
    
    metrics = [srcc, krcc, plcc, rmse];
end

function y = logistic_func(beta, x)
    % 四参数Logistic函数
    y = beta(2) + (beta(1) - beta(2)) ./ (1 + exp(-(x - beta(3)) ./ abs(beta(4))));
end