Opencv图像偏色检测

**

1. 偏色检测公式

---

**

图像的偏色不仅与图像色度的平均值有直接关系，还与图像的色度分布特性有关。如果在 a - b色度坐标平面上的二维直方图中色度分布基本上为单峰值，或者分布较为集中，而色度平均值又较大时，一般都存在偏色，而且色度平均值越大，偏色越严重。因此引入等效圆的概念，采用图像平均色度D和色度中心距M的比值，即偏色因子K来衡量图像的偏色程度。其计算方法如下式：

---

以上摘自论文**《基于图像分析的偏色检测及颜色校正方法》——徐晓昭，蔡轶珩等。**
但是在实际应用中，公式（3）去掉平方可更好的指示图像是否偏色：

**

2.RGB颜色空间转Lab颜色空间

**
颜色转换原理见：颜色空间系列2: RGB和CIELAB颜色空间的转换及优化算法
利用opencv实现代码为：

void RGB2LAB(Mat& rgb, Mat& Lab)
{
	Mat XYZ(rgb.size(), rgb.type());
	Mat_<Vec3b>::iterator begainRGB = rgb.begin<Vec3b>();
	Mat_<Vec3b>::iterator endRGB = rgb.end<Vec3b>();
	Mat_<Vec3b>::iterator begainXYZ = XYZ.begin<Vec3b>();
	int shift = 22;
	for (; begainRGB != endRGB; begainRGB++, begainXYZ++)
	{
		(*begainXYZ)[0] = ((*begainRGB)[0] * 199049 + (*begainRGB)[1] * 394494 + (*begainRGB)[2] * 455033 + 524288) >> (shift-2);
		(*begainXYZ)[1] = ((*begainRGB)[0] * 75675 + (*begainRGB)[1] * 749900 + (*begainRGB)[2] * 223002 + 524288) >> (shift-2);
		(*begainXYZ)[2] = ((*begainRGB)[0] * 915161 + (*begainRGB)[1] * 114795 + (*begainRGB)[2] * 18621 + 524288) >> (shift-2);
	}

	int LabTab[1024];
	for (int i = 0; i < 1024; i++)
	{
		if (i>9)
			LabTab[i] = (int)(pow((float)i / 1020, 1.0F / 3) * (1 << shift) + 0.5);
		else
			LabTab[i] = (int)((29 * 29.0 * i / (6 * 6 * 3 * 1020) + 4.0 / 29) * (1 << shift) + 0.5);
	}
	const int ScaleLC = (int)(16 * 2.55 * (1 << shift) + 0.5);
	const int ScaleLT = (int)(116 * 2.55 + 0.5);
	const int HalfShiftValue = 524288;
	begainXYZ = XYZ.begin<Vec3b>();
	Mat_<Vec3b>::iterator endXYZ = XYZ.end<Vec3b>();
	Lab.create(rgb.size(),rgb.type());
	Mat_<Vec3b>::iterator begainLab = Lab.begin<Vec3b>();
	for (; begainXYZ != endXYZ; begainXYZ++, begainLab++)
	{
		int X = LabTab[(*begainXYZ)[0]];
		int Y = LabTab[(*begainXYZ)[1]];
		int Z = LabTab[(*begainXYZ)[2]];
		int L = ((ScaleLT * Y - ScaleLC + HalfShiftValue) >> shift);
		int A = ((500 * (X - Y) + HalfShiftValue) >> shift) + 128;
		int B = ((200 * (Y - Z) + HalfShiftValue) >> shift) + 128;
		(*begainLab)[0] = L;
		(*begainLab)[1] = A;
		(*begainLab)[2] = B;
	}
}

3.偏色检测算法实现

根据偏色检测公式，opencv实现过程为：

float colorCheck(const Mat& imgLab)
{
	Mat_<Vec3b>::const_iterator begainIt = imgLab.begin<Vec3b>();
	Mat_<Vec3b>::const_iterator endIt = imgLab.end<Vec3b>();
	float aSum = 0;
	float bSum = 0;
	for (; begainIt != endIt; begainIt++)
	{
		aSum += (*begainIt)[1];
		bSum += (*begainIt)[2];
	}
	int MN = imgLab.cols*imgLab.rows;
	double Da = aSum / MN - 128; // 必须归一化到[-128，,127]范围内	
	double Db = bSum / MN - 128;

	//平均色度
	double D = sqrt(Da*Da+Db*Db);

	begainIt = imgLab.begin<Vec3b>();
	double Ma = 0;
	double Mb = 0;
	for (; begainIt != endIt; begainIt++)
	{
		Ma += abs((*begainIt)[1]-128 - Da);
		Mb += abs((*begainIt)[2]-128 - Db);
	}
	Ma = Ma / MN;
	Mb = Mb / MN;
	//色度中心距
	double M = sqrt(Ma*Ma + Mb*Mb);
	//偏色因子
	float K = (float)(D / M);
	return K;
}

综合来说，k值不大于1.5我们可以认为其整体图像偏色的可能性不大，当然这个值取多少可能还是需要和实际情况结合的。

---