opencv 证件照背景替换-KMeans

一、步骤说明

使用了KMeans图像分割，也使用了GMM高斯混合算法，但是感觉KMeans的效果好点。
整体的步骤：

1. 数据组装；
2. KMeans分割；
3. 背景去除；
4. 遮罩生成；
5. 遮罩模糊，（视情况而定，可以不做）；
6. 通道混合输出；
7. 完成。

二、代码实例

#include 
#include 
using namespace cv;
using namespace std;
using namespace ml;

const char* INPUT = "input image";

Mat mat_to_samples(Mat& image)
{
	int w = image.cols;
	int h = image.rows;
	int samplecount = w * h;
	int dims = image.channels();
	Mat points(samplecount, dims, CV_32F, Scalar(10));

	int index = 0;
	for (int row = 0; row < h; row++)
	{
		for (int col = 0; col < w; col++)
		{
			index = row * w + col;
			Vec3b bgr = image.at(row, col);
			points.at(index, 0) = static_cast(bgr[0]);
			points.at(index, 1) = static_cast(bgr[1]);
			points.at(index, 2) = static_cast(bgr[2]);
		}
	}
	return points;
}

int main()
{
	Mat src = imread("D:/source/images/zjz1.png");
	if (src.empty())
	{
		puts("read image error");
		system("pause");
		return -1;
	}
	imshow(INPUT, src);

	// 组装数据
	Mat points = mat_to_samples(src);

	// 运行 KMeans 感觉要比GMM分割的效果要好
	int numCluster = 4;
	Mat labels;
	Mat centers;
	TermCriteria citeria = TermCriteria(TermCriteria::EPS + TermCriteria::COUNT, 10, 0.1);
	kmeans(points, numCluster, labels, citeria, 3, KMEANS_PP_CENTERS, centers);

	// 使用GMM 分类算法
	//Ptr em_model = EM::create();
	//em_model->setClustersNumber(numCluster);
	//em_model->setCovarianceMatrixType(EM::COV_MAT_SPHERICAL);
	//em_model->setTermCriteria(TermCriteria(TermCriteria::EPS + TermCriteria::COUNT, 100, 0.1));
	//em_model->trainEM(points, noArray(), labels, noArray());

	// 去背景 + 遮罩生成
	Mat mask(src.size(), CV_8UC1); // 遮罩层
	int index = src.rows * 2 + 2;	// 取左上角的（2,2）坐标为背景颜色
	int cindex = labels.at(index, 0);
	int height = src.rows;
	int width = src.cols;
	Mat dst;
	src.copyTo(dst);

	for (int  row = 0; row < height; row++)
	{
		for (int col = 0; col < width; col++)
		{
			index = row * width + col;
			int label = labels.at(index, 0);
			if (label == cindex)
			{
				dst.at(row, col)[0] = 0; // 背景
				dst.at(row, col)[1] = 0; // 背景
				dst.at(row, col)[2] = 0; // 背景
				mask.at(row, col) = 0;
			}
			else
				mask.at(row, col) = 255; // 前景

		}
	}
	imshow("mask", mask);
	imshow("KMeans", dst);

	// 腐蚀 + 高斯模糊
	Mat k = getStructuringElement(MORPH_RECT, Size(3, 3), Point(-1, -1));
	erode(mask, mask, k);
	imshow("erode mask", mask);
	GaussianBlur(mask, mask, Size(3, 3), 0, 0); // 这里的核心是（3,3）一定要和腐蚀的核心一样，不然效果不好
	imshow("Gaussianblur mask", mask);

	// 通道混合 :获取权重，然后混合图像， 使得图像边缘过度自然
	RNG rng(12345);
	Vec3b color;
	color[0] = rng.uniform(0, 255);
	color[1] = rng.uniform(0, 255);
	color[2] = rng.uniform(0, 255);
	Mat result(src.size(), src.type());

	double w = 0.0; // 权重
	int b = 0, g = 0, r = 0; // 通道混合
	int b1 = 0, g1 = 0, r1 = 0; // 前景
	int b2 = 0, g2 = 0, r2 = 0; // 背景

	for (int row = 0; row < height; row++)
	{
		for (int col = 0; col < width; col++)
		{
			int m = mask.at(row, col);
			if (m == 255) // 前景直接复制原图像。
			{
				result.at(row, col) = src.at(row, col);
			}
			else if (m == 0) // 背景
			{
				result.at(row, col) = color;
			}
			else
			{
				w = m / 255;
				b1 = src.at(row, col)[0]; // 前景
				g1 = src.at(row, col)[1];
				r1 = src.at(row, col)[2];

				b2 = color[0]; // 背景
				g2 = color[1];
				r2 = color[2];

				b = b1 * w + b2 * (1.0 - w);
				g = g1 * w + g2 * (1.0 - w);
				r = r1 * w + r2 * (1.0 - w);

				result.at(row, col)[0] = b;
				result.at(row, col)[1] = g;
				result.at(row, col)[2] = r;
			}
		}
	}
	imshow("背景替换", result);

	waitKey(0);
	return 0;
}

三、结果展示