#pragma once
#include <iostream>
#include <ctime>

#include "opencv2/opencv.hpp"

class QuantKMeans
{
public:
    int     m_alpha_thres;
    cv::Mat m_flatten_labels;
    cv::Mat m_centers8UC3;
    QuantKMeans(int alpha_thres = 10) : m_alpha_thres(alpha_thres) {}
    void apply(cv::Mat &ori_image, cv::Mat &new_image, int num_cluster, int color_space)
    {
        cv::Mat image;
        convert_color_space(ori_image, image, color_space);
        cv::Mat flatten_image = flatten(image);

        apply(flatten_image, num_cluster, color_space);
        replace_centers(ori_image, new_image);
    }
    void apply_aplha(cv::Mat &ori_image, cv::Mat &new_image, int num_cluster, int color_space)
    {
        // cout << " *** DoAlpha *** " << endl;
        cv::Mat flatten_image8UC3 = flatten_alpha(ori_image);
        cv::Mat image8UC3;
        convert_color_space(flatten_image8UC3, image8UC3, color_space);
        cv::Mat image32FC3(image8UC3.rows, 1, CV_32FC3);
        for (int i = 0; i < image8UC3.rows; i++) image32FC3.at<cv::Vec3f>(i, 0) = image8UC3.at<cv::Vec3b>(i, 0);

        apply(image32FC3, num_cluster, color_space);
        repalce_centers_aplha(ori_image, new_image);
    }
    void apply(cv::Mat &flatten_image, int num_cluster, int color_space)
    {
        cv::Mat centers32FC3;
        num_cluster = fmin(flatten_image.rows, num_cluster);
        kmeans(flatten_image, num_cluster, this->m_flatten_labels, cv::TermCriteria(cv::TermCriteria::EPS + cv::TermCriteria::COUNT, 300, 0.5), 3, cv::KMEANS_PP_CENTERS,
               centers32FC3);
        this->m_centers8UC3 = cv::Mat(num_cluster, 1, CV_8UC3);
        for (int i = 0; i < num_cluster; i++) this->m_centers8UC3.at<cv::Vec3b>(i) = centers32FC3.at<cv::Vec3f>(i);

        convert_color_space(this->m_centers8UC3, this->m_centers8UC3, color_space, true);
    }

    void apply(const std::vector<std::array<float, 4>> &ori_colors,
               std::vector<std::array<float, 4>> &      cluster_results,
               std::vector<int> &                       labels,
               int                                      num_cluster = -1,
               int                                      color_space = 2)
    {
        // 0~255
        cv::Mat flatten_image8UC3 = flatten_vector(ori_colors);

        cv::Mat image8UC3;
        convert_color_space(flatten_image8UC3, image8UC3, color_space);

        cv::Mat image32FC3(image8UC3.rows, 1, CV_32FC3);
        for (int i = 0; i < image8UC3.rows; i++) image32FC3.at<cv::Vec3f>(i, 0) = image8UC3.at<cv::Vec3b>(i, 0);

        int best_cluster = 1, cur_score = 0, best_score = 100;
        int max_cluster = ori_colors.size();
        num_cluster     = fmin(num_cluster, max_cluster);
        if (num_cluster < 1) {
            cur_score  = kmeans(image32FC3, 1, this->m_flatten_labels, cv::TermCriteria(cv::TermCriteria::EPS + cv::TermCriteria::COUNT, 300, 0.5), 3, cv::KMEANS_PP_CENTERS);
            best_score = cur_score;
            for (int cur_cluster = 2; cur_cluster < 16; cur_cluster++) {
                if (cur_cluster > max_cluster) break;
                cur_score    = kmeans(image32FC3, cur_cluster, this->m_flatten_labels, cv::TermCriteria(cv::TermCriteria::EPS + cv::TermCriteria::COUNT, 300, 0.5), 3,
                                   cv::KMEANS_PP_CENTERS);
                best_cluster = cur_score < best_score ? cur_cluster : best_cluster;
                best_score   = cur_score < best_score ? cur_score : best_score;
            }
        } else
            best_cluster = num_cluster;

        cv::Mat centers32FC3;
        kmeans(image32FC3, best_cluster, this->m_flatten_labels, cv::TermCriteria(cv::TermCriteria::EPS + cv::TermCriteria::COUNT, 300, 0.5), 3, cv::KMEANS_PP_CENTERS,
               centers32FC3);
        this->m_centers8UC3 = cv::Mat(best_cluster, 1, CV_8UC3);
        for (int i = 0; i < best_cluster; i++) this->m_centers8UC3.at<cv::Vec3b>(i) = centers32FC3.at<cv::Vec3f>(i);

        convert_color_space(this->m_centers8UC3, this->m_centers8UC3, color_space, true);

        cluster_results.clear();
        labels.clear();
        for (int i = 0; i < ori_colors.size(); i++)
            labels.emplace_back(this->m_flatten_labels.at<int>(i, 0));
        for (int i = 0; i < best_cluster; i++) {
            cv::Vec3f center = this->m_centers8UC3.at<cv::Vec3b>(i, 0);
            cluster_results.emplace_back(std::array<float, 4>{center[0] / 255.f, center[1] / 255.f, center[2] / 255.f, 1.f});
        }
    }

    void replace_centers(cv::Mat &ori_image, cv::Mat &new_image)
    {
        for (int i = 0; i < ori_image.rows; i++) {
            for (int j = 0; j < ori_image.cols; j++) {
                int       idx                 = this->m_flatten_labels.at<int>(i * ori_image.cols + j, 0);
                cv::Vec3b pixel               = this->m_centers8UC3.at<cv::Vec3b>(idx);
                new_image.at<cv::Vec3b>(i, j) = pixel;
            }
        }
    }
    void repalce_centers_aplha(cv::Mat &ori_image, cv::Mat &new_image)
    {
        int       cnt = 0;
        int       idx;
        cv::Vec3b center;
        for (int i = 0; i < ori_image.rows; i++) {
            for (int j = 0; j < ori_image.cols; j++) {
                cv::Vec4b pixel = ori_image.at<cv::Vec4b>(i, j);
                if ((int) pixel[3] < this->m_alpha_thres)
                    new_image.at<cv::Vec4b>(i, j) = pixel;
                else {
                    idx                           = this->m_flatten_labels.at<int>(cnt++, 0);
                    center                        = this->m_centers8UC3.at<cv::Vec3b>(idx);
                    new_image.at<cv::Vec4b>(i, j) = cv::Vec4b(center[0], center[1], center[2], pixel[3]);
                }
            }
        }
    }

    void convert_color_space(cv::Mat &ori_image, cv::Mat &image, int color_space, bool reverse = false)
    {
        switch (color_space) {
        case 0: image = ori_image; break;
        case 1:
            if (reverse)
                cvtColor(ori_image, image, cv::COLOR_HSV2BGR);
            else
                cvtColor(ori_image, image, cv::COLOR_BGR2HSV);
            break;
        case 2:
            if (reverse)
                cvtColor(ori_image, image, cv::COLOR_Lab2BGR);
            else
                cvtColor(ori_image, image, cv::COLOR_BGR2Lab);
            break;
        default: break;
        }
    }

    cv::Mat flatten(cv::Mat &image)
    {
        int     num_pixels = image.rows * image.cols;
        cv::Mat img(num_pixels, 1, CV_32FC3);
        for (int i = 0; i < image.rows; i++) {
            for (int j = 0; j < image.cols; j++) {
                cv::Vec3f pixel                          = image.at<cv::Vec3b>(i, j);
                img.at<cv::Vec3f>(i * image.cols + j, 0) = pixel;
            }
        }
        return img;
    }
    cv::Mat flatten_alpha(cv::Mat &image)
    {
        int num_pixels = image.rows * image.cols;
        for (int i = 0; i < image.rows; i++)
            for (int j = 0; j < image.cols; j++) {
                cv::Vec4b pixel = image.at<cv::Vec4b>(i, j);
                if ((int) pixel[3] < this->m_alpha_thres) num_pixels--;
            }

        cv::Mat img(num_pixels, 1, CV_8UC3);
        int     cnt = 0;
        for (int i = 0; i < image.rows; i++) {
            for (int j = 0; j < image.cols; j++) {
                cv::Vec4b pixel = image.at<cv::Vec4b>(i, j);
                if ((int) pixel[3] >= this->m_alpha_thres) img.at<cv::Vec3b>(cnt++, 0) = cv::Vec3b(pixel[0], pixel[1], pixel[2]);
            }
        }
        return img;
    }
    cv::Mat flatten_vector(const std::vector<std::array<float, 4>> &ori_colors)
    {
        int num_pixels = ori_colors.size();

        cv::Mat image8UC3(num_pixels, 1, CV_8UC3);
        for (int i = 0; i < num_pixels; i++) {
            std::array<float, 4> pixel    = ori_colors[i];
            image8UC3.at<cv::Vec3b>(i, 0) = cv::Vec3b((int) (pixel[0] * 255.f), (int) (pixel[1] * 255.f), (int) (pixel[2] * 255.f));
        }
        return image8UC3;
    }
};