OpenCV教程（14）问题思考---- 深度学习中为什么普遍使用BGR而不用RGB？

Table of Contents

1. 发现问题

现象：

2. 提出问题

问题：为什么使用BRG描述图像，而非RGB通道？

3. 合理假设

4. 求解

stackoverflow解答

1. 发现问题

计算机视觉中普遍用BRG描述图像，而非RGB通道？

现象：

查看OpenCV ImageProcessing 模块内对色彩空间的定义和使用，官方文档对imgproc的描述如下：

#include "opencv2/core.hpp"
#include "opencv2/imgproc/imgproc_c.h"
Classes
class   cv::CLAHEBase class for Contrast Limited Adaptive Histogram Equalization. : More...
class   cv::GeneralizedHoughfinds arbitrary template in the grayscale image using Generalized Hough Transform
class   cv::GeneralizedHoughBallard
class   cv::GeneralizedHoughGuil
class   cv::LineIteratorLine iterator. More...
class   cv::LineSegmentDetectorLine segment detector class. More...
struct      cv::Subdiv2D::QuadEdge
class   cv::Subdiv2D
struct      cv::Subdiv2D::VertexNamespacescvMacros
#define     CV_RGB(r, g, b)   cv::Scalar((b), (g), (r), 0)

imgproc模块中的色彩转换函数声明：

enum     cv::ColorConversionCodes { cv::COLOR_BGR2BGRA = 0, cv::COLOR_RGB2RGBA = COLOR_BGR2BGRA, cv::COLOR_BGRA2BGR = 1, cv::COLOR_RGBA2RGB = COLOR_BGRA2BGR, cv::COLOR_BGR2RGBA = 2, cv::COLOR_RGB2BGRA = COLOR_BGR2RGBA, cv::COLOR_RGBA2BGR = 3, cv::COLOR_BGRA2RGB = COLOR_RGBA2BGR, cv::COLOR_BGR2RGB = 4, cv::COLOR_RGB2BGR = COLOR_BGR2RGB, cv::COLOR_BGRA2RGBA = 5, cv::COLOR_RGBA2BGRA = COLOR_BGRA2RGBA, cv::COLOR_BGR2GRAY = 6, cv::COLOR_RGB2GRAY = 7, cv::COLOR_GRAY2BGR = 8, cv::COLOR_GRAY2RGB = COLOR_GRAY2BGR, cv::COLOR_GRAY2BGRA = 9, cv::COLOR_GRAY2RGBA = COLOR_GRAY2BGRA, cv::COLOR_BGRA2GRAY = 10, cv::COLOR_RGBA2GRAY = 11, cv::COLOR_BGR2BGR565 = 12, cv::COLOR_RGB2BGR565 = 13, cv::COLOR_BGR5652BGR = 14, cv::COLOR_BGR5652RGB = 15, cv::COLOR_BGRA2BGR565 = 16, cv::COLOR_RGBA2BGR565 = 17, cv::COLOR_BGR5652BGRA = 18, cv::COLOR_BGR5652RGBA = 19, cv::COLOR_GRAY2BGR565 = 20, cv::COLOR_BGR5652GRAY = 21, cv::COLOR_BGR2BGR555 = 22, cv::COLOR_RGB2BGR555 = 23, cv::COLOR_BGR5552BGR = 24, cv::COLOR_BGR5552RGB = 25, cv::COLOR_BGRA2BGR555 = 26, cv::COLOR_RGBA2BGR555 = 27, cv::COLOR_BGR5552BGRA = 28, cv::COLOR_BGR5552RGBA = 29, cv::COLOR_GRAY2BGR555 = 30, cv::COLOR_BGR5552GRAY = 31, cv::COLOR_BGR2XYZ = 32, cv::COLOR_RGB2XYZ = 33, cv::COLOR_XYZ2BGR = 34, cv::COLOR_XYZ2RGB = 35, cv::COLOR_BGR2YCrCb = 36, cv::COLOR_RGB2YCrCb = 37, cv::COLOR_YCrCb2BGR = 38, cv::COLOR_YCrCb2RGB = 39, cv::COLOR_BGR2HSV = 40, cv::COLOR_RGB2HSV = 41, cv::COLOR_BGR2Lab = 44, cv::COLOR_RGB2Lab = 45, cv::COLOR_BGR2Luv = 50, cv::COLOR_RGB2Luv = 51, cv::COLOR_BGR2HLS = 52, cv::COLOR_RGB2HLS = 53, cv::COLOR_HSV2BGR = 54, cv::COLOR_HSV2RGB = 55, cv::COLOR_Lab2BGR = 56, cv::COLOR_Lab2RGB = 57, cv::COLOR_Luv2BGR = 58, cv::COLOR_Luv2RGB = 59, cv::COLOR_HLS2BGR = 60, cv::COLOR_HLS2RGB = 61, cv::COLOR_BGR2HSV_FULL = 66, cv::COLOR_RGB2HSV_FULL = 67, cv::COLOR_BGR2HLS_FULL = 68, cv::COLOR_RGB2HLS_FULL = 69, cv::COLOR_HSV2BGR_FULL = 70, cv::COLOR_HSV2RGB_FULL = 71, cv::COLOR_HLS2BGR_FULL = 72, cv::COLOR_HLS2RGB_FULL = 73, cv::COLOR_LBGR2Lab = 74, cv::COLOR_LRGB2Lab = 75, cv::COLOR_LBGR2Luv = 76, cv::COLOR_LRGB2Luv = 77, cv::COLOR_Lab2LBGR = 78, cv::COLOR_Lab2LRGB = 79, cv::COLOR_Luv2LBGR = 80, cv::COLOR_Luv2LRGB = 81, cv::COLOR_BGR2YUV = 82, cv::COLOR_RGB2YUV = 83, cv::COLOR_YUV2BGR = 84, cv::COLOR_YUV2RGB = 85, cv::COLOR_YUV2RGB_NV12 = 90, cv::COLOR_YUV2BGR_NV12 = 91, cv::COLOR_YUV2RGB_NV21 = 92, cv::COLOR_YUV2BGR_NV21 = 93, cv::COLOR_YUV420sp2RGB = COLOR_YUV2RGB_NV21, cv::COLOR_YUV420sp2BGR = COLOR_YUV2BGR_NV21, cv::COLOR_YUV2RGBA_NV12 = 94, cv::COLOR_YUV2BGRA_NV12 = 95, cv::COLOR_YUV2RGBA_NV21 = 96, cv::COLOR_YUV2BGRA_NV21 = 97, cv::COLOR_YUV420sp2RGBA = COLOR_YUV2RGBA_NV21, cv::COLOR_YUV420sp2BGRA = COLOR_YUV2BGRA_NV21, cv::COLOR_YUV2RGB_YV12 = 98, cv::COLOR_YUV2BGR_YV12 = 99, cv::COLOR_YUV2RGB_IYUV = 100, cv::COLOR_YUV2BGR_IYUV = 101, cv::COLOR_YUV2RGB_I420 = COLOR_YUV2RGB_IYUV, cv::COLOR_YUV2BGR_I420 = COLOR_YUV2BGR_IYUV, cv::COLOR_YUV420p2RGB = COLOR_YUV2RGB_YV12, cv::COLOR_YUV420p2BGR = COLOR_YUV2BGR_YV12, cv::COLOR_YUV2RGBA_YV12 = 102, cv::COLOR_YUV2BGRA_YV12 = 103, cv::COLOR_YUV2RGBA_IYUV = 104, cv::COLOR_YUV2BGRA_IYUV = 105, cv::COLOR_YUV2RGBA_I420 = COLOR_YUV2RGBA_IYUV, cv::COLOR_YUV2BGRA_I420 = COLOR_YUV2BGRA_IYUV, cv::COLOR_YUV420p2RGBA = COLOR_YUV2RGBA_YV12, cv::COLOR_YUV420p2BGRA = COLOR_YUV2BGRA_YV12, cv::COLOR_YUV2GRAY_420 = 106, cv::COLOR_YUV2GRAY_NV21 = COLOR_YUV2GRAY_420, cv::COLOR_YUV2GRAY_NV12 = COLOR_YUV2GRAY_420, cv::COLOR_YUV2GRAY_YV12 = COLOR_YUV2GRAY_420, cv::COLOR_YUV2GRAY_IYUV = COLOR_YUV2GRAY_420, cv::COLOR_YUV2GRAY_I420 = COLOR_YUV2GRAY_420, cv::COLOR_YUV420sp2GRAY = COLOR_YUV2GRAY_420, cv::COLOR_YUV420p2GRAY = COLOR_YUV2GRAY_420, cv::COLOR_YUV2RGB_UYVY = 107, cv::COLOR_YUV2BGR_UYVY = 108, cv::COLOR_YUV2RGB_Y422 = COLOR_YUV2RGB_UYVY, cv::COLOR_YUV2BGR_Y422 = COLOR_YUV2BGR_UYVY, cv::COLOR_YUV2RGB_UYNV = COLOR_YUV2RGB_UYVY, cv::COLOR_YUV2BGR_UYNV = COLOR_YUV2BGR_UYVY, cv::COLOR_YUV2RGBA_UYVY = 111, cv::COLOR_YUV2BGRA_UYVY = 112, cv::COLOR_YUV2RGBA_Y422 = COLOR_YUV2RGBA_UYVY, cv::COLOR_YUV2BGRA_Y422 = COLOR_YUV2BGRA_UYVY, cv::COLOR_YUV2RGBA_UYNV = COLOR_YUV2RGBA_UYVY, cv::COLOR_YUV2BGRA_UYNV = COLOR_YUV2BGRA_UYVY, cv::COLOR_YUV2RGB_YUY2 = 115, cv::COLOR_YUV2BGR_YUY2 = 116, cv::COLOR_YUV2RGB_YVYU = 117, cv::COLOR_YUV2BGR_YVYU = 118, cv::COLOR_YUV2RGB_YUYV = COLOR_YUV2RGB_YUY2, cv::COLOR_YUV2BGR_YUYV = COLOR_YUV2BGR_YUY2, cv::COLOR_YUV2RGB_YUNV = COLOR_YUV2RGB_YUY2, cv::COLOR_YUV2BGR_YUNV = COLOR_YUV2BGR_YUY2, cv::COLOR_YUV2RGBA_YUY2 = 119, cv::COLOR_YUV2BGRA_YUY2 = 120, cv::COLOR_YUV2RGBA_YVYU = 121, cv::COLOR_YUV2BGRA_YVYU = 122, cv::COLOR_YUV2RGBA_YUYV = COLOR_YUV2RGBA_YUY2, cv::COLOR_YUV2BGRA_YUYV = COLOR_YUV2BGRA_YUY2, cv::COLOR_YUV2RGBA_YUNV = COLOR_YUV2RGBA_YUY2, cv::COLOR_YUV2BGRA_YUNV = COLOR_YUV2BGRA_YUY2, cv::COLOR_YUV2GRAY_UYVY = 123, cv::COLOR_YUV2GRAY_YUY2 = 124, cv::COLOR_YUV2GRAY_Y422 = COLOR_YUV2GRAY_UYVY, cv::COLOR_YUV2GRAY_UYNV = COLOR_YUV2GRAY_UYVY, cv::COLOR_YUV2GRAY_YVYU = COLOR_YUV2GRAY_YUY2, cv::COLOR_YUV2GRAY_YUYV = COLOR_YUV2GRAY_YUY2, cv::COLOR_YUV2GRAY_YUNV = COLOR_YUV2GRAY_YUY2, cv::COLOR_RGBA2mRGBA = 125, cv::COLOR_mRGBA2RGBA = 126, cv::COLOR_RGB2YUV_I420 = 127, cv::COLOR_BGR2YUV_I420 = 128, cv::COLOR_RGB2YUV_IYUV = COLOR_RGB2YUV_I420, cv::COLOR_BGR2YUV_IYUV = COLOR_BGR2YUV_I420, cv::COLOR_RGBA2YUV_I420 = 129, cv::COLOR_BGRA2YUV_I420 = 130, cv::COLOR_RGBA2YUV_IYUV = COLOR_RGBA2YUV_I420, cv::COLOR_BGRA2YUV_IYUV = COLOR_BGRA2YUV_I420, cv::COLOR_RGB2YUV_YV12 = 131, cv::COLOR_BGR2YUV_YV12 = 132, cv::COLOR_RGBA2YUV_YV12 = 133, cv::COLOR_BGRA2YUV_YV12 = 134, cv::COLOR_BayerBG2BGR = 46, cv::COLOR_BayerGB2BGR = 47, cv::COLOR_BayerRG2BGR = 48, cv::COLOR_BayerGR2BGR = 49, cv::COLOR_BayerBG2RGB = COLOR_BayerRG2BGR, cv::COLOR_BayerGB2RGB = COLOR_BayerGR2BGR, cv::COLOR_BayerRG2RGB = COLOR_BayerBG2BGR, cv::COLOR_BayerGR2RGB = COLOR_BayerGB2BGR, cv::COLOR_BayerBG2GRAY = 86, cv::COLOR_BayerGB2GRAY = 87, cv::COLOR_BayerRG2GRAY = 88, cv::COLOR_BayerGR2GRAY = 89, cv::COLOR_BayerBG2BGR_VNG = 62, cv::COLOR_BayerGB2BGR_VNG = 63, cv::COLOR_BayerRG2BGR_VNG = 64, cv::COLOR_BayerGR2BGR_VNG = 65, cv::COLOR_BayerBG2RGB_VNG = COLOR_BayerRG2BGR_VNG, cv::COLOR_BayerGB2RGB_VNG = COLOR_BayerGR2BGR_VNG, cv::COLOR_BayerRG2RGB_VNG = COLOR_BayerBG2BGR_VNG, cv::COLOR_BayerGR2RGB_VNG = COLOR_BayerGB2BGR_VNG, cv::COLOR_BayerBG2BGR_EA = 135, cv::COLOR_BayerGB2BGR_EA = 136, cv::COLOR_BayerRG2BGR_EA = 137, cv::COLOR_BayerGR2BGR_EA = 138, cv::COLOR_BayerBG2RGB_EA = COLOR_BayerRG2BGR_EA, cv::COLOR_BayerGB2RGB_EA = COLOR_BayerGR2BGR_EA, cv::COLOR_BayerRG2RGB_EA = COLOR_BayerBG2BGR_EA, cv::COLOR_BayerGR2RGB_EA = COLOR_BayerGB2BGR_EA, cv::COLOR_BayerBG2BGRA = 139, cv::COLOR_BayerGB2BGRA = 140, cv::COLOR_BayerRG2BGRA = 141, cv::COLOR_BayerGR2BGRA = 142, cv::COLOR_BayerBG2RGBA = COLOR_BayerRG2BGRA, cv::COLOR_BayerGB2RGBA = COLOR_BayerGR2BGRA, cv::COLOR_BayerRG2RGBA = COLOR_BayerBG2BGRA, cv::COLOR_BayerGR2RGBA = COLOR_BayerGB2BGRA, cv::COLOR_COLORCVT_MAX = 143
}

可以看到OpenCV提供了cv::COLOR_RGB2BGR()和 cv::COLOR_BGR2RGB ()等多个函数，用于色彩空间转换，而OpenCV默认使用的恰恰是BGR色彩空间。

2. 提出问题

问题：为什么使用BRG描述图像，而非RGB通道？

3. 合理假设

OpenCV默认通道为BGR，可能是基于某种硬件层面的原因。

4. 求解

stackoverflow解答

RGB is a byte-order. But a deliberate implementation choice of most vanilla Graphics libraries is that they treat colours as unsigned 32-bit integers internally, with the three (or four, as alpha is typically included) components packed into the integer.

On a little-endian machine (such as x86) the integer 0x01020304 will actually be stored in memory as 0x04030201. And thus 0x00BBGGRR will be stored as 0xRRGGBB00!

So the term BGR (and BGRA etc) is a leaky abstraction where the graphics library is explaining how the integer is logically ordered, so as to make your code that is directly accessing the colour components individually more readable.

Remember that bitmaps are usually accessed by more parts of the hardware than your processor, and the endian that is specified by, say, conventional display adapters, is not necessarily the same as the endian of your CPU. At the level of manipulating the channels in the pixel its no problem for a CPU to extract the fields whatever their order; its purely a programmer understanding the labelling thing.

翻译：

RGB是字节顺序。但是大多数vanilla图形库的有意实现选择是它们在内部将颜色视为无符号32位整数，其中三个（或四个，通常包括alpha）组件被打包到整数中。

在小端机器（例如x86）上，整数0x01020304实际上将作为0x04030201存储在内存中。因此0x00BBGGRR将存储为0xRRGGBB00！

所以术语BGR（和BGRA等）是一个漏洞抽象，其中图形库解释整数是如何按逻辑排序的，以便使直接访问颜色组件的代码更具可读性。

请记住，位图通常由硬件的更多部分访问，而不是处理器，并且由传统显示适配器指定的字节序不一定与CPU的字节序相同。在操纵像素中的通道的级别上，CPU无论其顺序如何都可以提取字段;它纯粹是程序员理解标签的东西。

知乎回答：

问题：为什么深度学习中普遍用BRG描述图像，而非RGB通道？

答1：因为caffe，作为最早最流行的一批库的代表，用了opencv，而opencv默认通道是bgr的。这是opencv的入门大坑之一，bgr是个历史遗留问题，为了兼容早年的某些硬件。其实你自己训练完全可以用rgb，新库也基本没了bgr还是rgb这个问题，就是切换下顺序。但如果你要用一些老的训练好的模型，就得兼容老模型的bgr。

答2：因为OpenCV默认通道为BGR，至于为什么用BGR，可能是因为当时比较流行BGR，以至于后来RGB变为主流后，也不方便改了，就一直沿用BGR。而caffe又是用了opencv的，所以没办法。智能外部转换一下。