【OpenCV】Switching Eds: Face swapping

Switching Eds: Face swapping with Python, dlib, and OpenCV

Face Swap using OpenCV ( C++ / Python )

肤色变换：

void specifiyHistogram(const cv::Mat source_image, cv::Mat target_image, cv::Mat mask)
{
	int source_hist_int[3][256];
	int target_hist_int[3][256];
	float source_histogram[3][256];
	float target_histogram[3][256];
	std::memset(source_hist_int, 0, sizeof(int) * 3 * 256);
	std::memset(target_hist_int, 0, sizeof(int) * 3 * 256);

	for (size_t i = 0; i < mask.rows; i++)
	{
		auto current_mask_pixel = mask.row(i).data;
		auto current_source_pixel = source_image.row(i).data;
		auto current_target_pixel = target_image.row(i).data;

		for (size_t j = 0; j < mask.cols; j++)
		{
			if (*current_mask_pixel != 0) {
				source_hist_int[0][*current_source_pixel]++;
				source_hist_int[1][*(current_source_pixel + 1)]++;
				source_hist_int[2][*(current_source_pixel + 2)]++;

				target_hist_int[0][*current_target_pixel]++;
				target_hist_int[1][*(current_target_pixel + 1)]++;
				target_hist_int[2][*(current_target_pixel + 2)]++;
			}

			// Advance to next pixel
			current_source_pixel += 3;
			current_target_pixel += 3;
			current_mask_pixel++;
		}
	}

	// Calc CDF
	for (size_t i = 1; i < 256; i++)
	{
		source_hist_int[0][i] += source_hist_int[0][i - 1];
		source_hist_int[1][i] += source_hist_int[1][i - 1];
		source_hist_int[2][i] += source_hist_int[2][i - 1];

		target_hist_int[0][i] += target_hist_int[0][i - 1];
		target_hist_int[1][i] += target_hist_int[1][i - 1];
		target_hist_int[2][i] += target_hist_int[2][i - 1];
	}

	// Normalize CDF
	for (size_t i = 0; i < 256; i++)
	{
		source_histogram[0][i] = (source_hist_int[0][i] ? (float)source_hist_int[0][i] / source_hist_int[0][255] : 0);
		source_histogram[1][i] = (source_hist_int[1][i] ? (float)source_hist_int[1][i] / source_hist_int[1][255] : 0);
		source_histogram[2][i] = (source_hist_int[2][i] ? (float)source_hist_int[2][i] / source_hist_int[2][255] : 0);

		target_histogram[0][i] = (target_hist_int[0][i] ? (float)target_hist_int[0][i] / target_hist_int[0][255] : 0);
		target_histogram[1][i] = (target_hist_int[1][i] ? (float)target_hist_int[1][i] / target_hist_int[1][255] : 0);
		target_histogram[2][i] = (target_hist_int[2][i] ? (float)target_hist_int[2][i] / target_hist_int[2][255] : 0);
	}

	// Create lookup table

	auto binary_search = [&](const float needle, const float haystack[]) -> uint8_t
	{
		uint8_t l = 0, r = 255, m;
		while (l < r)
		{
			m = (l + r) / 2;
			if (needle > haystack[m])
				l = m + 1;
			else
				r = m - 1;
		}
		// TODO check closest value
		return m;
	};

	uint8_t LUT[3][256];
	for (size_t i = 0; i < 256; i++)
	{
		LUT[0][i] = binary_search(target_histogram[0][i], source_histogram[0]);
		LUT[1][i] = binary_search(target_histogram[1][i], source_histogram[1]);
		LUT[2][i] = binary_search(target_histogram[2][i], source_histogram[2]);
	}

	// repaint pixels
	for (size_t i = 0; i < mask.rows; i++)
	{
		auto current_mask_pixel = mask.row(i).data;
		auto current_target_pixel = target_image.row(i).data;
		for (size_t j = 0; j < mask.cols; j++)
		{
			if (*current_mask_pixel != 0)
			{
				*current_target_pixel = LUT[0][*current_target_pixel];
				*(current_target_pixel + 1) = LUT[1][*(current_target_pixel + 1)];
				*(current_target_pixel + 2) = LUT[2][*(current_target_pixel + 2)];
			}

			// Advance to next pixel
			current_target_pixel += 3;
			current_mask_pixel++;
		}
	}
}