nv12, yuv420, yuv444转换关系

nv12, yuv420, yuv444转换关系

对于宽高为(w,h)的图像

• yuv444

  yuv444的排列顺序：[y, y, y, y, …, y, u, u,u,u,…u, v,v,v,v,…v], shape:(h, w, 3).
  乘以一个3*3矩阵可以转换为RGB, opencv已经提供了这种函数：cv2.cvtColor(yuv444, cv2.COLORYUV2RGB).

• yuv420

  排列顺序: [y, y, y, y, …, y, u, u, u, …u, v, v, v, …, v], shape:(h*3/2, w),
  其中前[:h, :]为y, [h : h * 3/2, :]为u和v，先u,再v. cvtColor转换flag：cv2.COLORYUV2RGB_I420.

• nv12

  排列顺序：[y,y,y,y,…,y, u,v,u,v,…], shape和yuv420一样：(h*3/2, w),
  前[:h, :]为y, 剩下部分u,v间隔排列。与yuv420唯一区别是uv顺序, cv2.COLORYUV2RGB_NV12.

  nv12toyuv转换方式：

  def convert_nv12_to_yuv444_naive(width, height, nv12_data):
      yuv444 = np.zeros((height, width,  3), nv12_data.dtype)
      yuv444[:, :, 0] = nv12_data[:width * height].reshape(height, width)
      uv_start = width * height
      import math
      for j in range(height):
          for i in range(width):
              uv_idx = uv_start + (width * math.floor(j / 2)) + (math.floor(i / 2))*2
              u = nv12_data[uv_idx]
              v = nv12_data[uv_idx+1]
              yuv444[j, i , 1:] = u, v
      return yuv444
  

  yuv444不够的uv要从nv12的uv上重复采样。

参考：

http://paulbourke.net/dataformats/nv12/