基于二维伽马函数的光照不均匀图像自适应校正算法

关于数字图像处理：本文介绍一种用于解决光照不均匀的图像自适应校正算法。光照不均匀其实是非常常见的一种状况，为了解决图像曝光不足，或者不均衡而提出来的解决方案之一，关于【Retiex解决方案请查看】。
论文原文: 点击查看

算法原理
论文使用了Retinex的多尺度高斯滤波求取「光照分量」，然后使用了「二维Gamma函数」针对原图的「HSV空间的V(亮度)分量」进行亮度改变，得到结果。

具体实现流程如下：

其中：

上面图像中 $\gamma$ 应该是：
$$\gamma ={\left(\frac{1}{2}\right)}^{\frac{m-I(x,y)}{m}}$$

Mat RGB2HSV(Mat src) {
	int row = src.rows;
	int col = src.cols;
	Mat dst(row, col, CV_32FC3);
	for (int i = 0; i < row; i++) {
		for (int j = 0; j < col; j++) {
			float b = src.at<Vec3b>(i, j)[0] / 255.0;
			float g = src.at<Vec3b>(i, j)[1] / 255.0;
			float r = src.at<Vec3b>(i, j)[2] / 255.0;
			float minn = min(r, min(g, b));
			float maxx = max(r, max(g, b));
			dst.at<Vec3f>(i, j)[2] = maxx; //V
			float delta = maxx - minn;
			float h, s;
			if (maxx != 0) {
				s = delta / maxx;
			}
			else {
				s = 0;
			}
			if (r == maxx) {
				h = (g - b) / delta;
			}
			else if (g == maxx) {
				h = 2 + (b - r) / delta;
			}
			else {
				h = 4 + (r - g) / delta;
			}
			h *= 60;
			if (h < 0)
				h += 360;
			dst.at<Vec3f>(i, j)[0] = h;
			dst.at<Vec3f>(i, j)[1] = s;
		}
	}
	return dst;
}

Mat HSV2RGB(Mat src) {
	int row = src.rows;
	int col = src.cols;
	Mat dst(row, col, CV_8UC3);
	float r, g, b, h, s, v;
	for (int i = 0; i < row; i++) {
		for (int j = 0; j < col; j++) {
			h = src.at<Vec3f>(i, j)[0];
			s = src.at<Vec3f>(i, j)[1];
			v = src.at<Vec3f>(i, j)[2];
			if (s == 0) {
				r = g = b = v;
			}
			else {
				h /= 60;
				int offset = floor(h);
				float f = h - offset;
				float p = v * (1 - s);
				float q = v * (1 - s * f);
				float t = v * (1 - s * (1 - f));
				switch (offset)
				{
				case 0: r = v; g = t; b = p; break;
				case 1: r = q; g = v; b = p; break;
				case 2: r = p; g = v; b = t; break;
				case 3: r = p; g = q; b = v; break;
				case 4: r = t; g = p; b = v; break;
				case 5: r = v; g = p; b = q; break;
				default:
					break;
				}
			}
			dst.at<Vec3b>(i, j)[0] = int(b * 255);
			dst.at<Vec3b>(i, j)[1] = int(g * 255);
			dst.at<Vec3b>(i, j)[2] = int(r * 255);
		}
	}
	return dst;
}

Mat work(Mat src) {
	int row = src.rows;
	int col = src.cols;
	Mat now = RGB2HSV(src);
	Mat H(row, col, CV_32FC1);
	Mat S(row, col, CV_32FC1);
	Mat V(row, col, CV_32FC1);
	for (int i = 0; i < row; i++) {
		for (int j = 0; j < col; j++) {
			H.at<float>(i, j) = now.at<Vec3f>(i, j)[0];
			S.at<float>(i, j) = now.at<Vec3f>(i, j)[1];
			V.at<float>(i, j) = now.at<Vec3f>(i, j)[2];
		}
	}
	int kernel_size = min(row, col);
	if (kernel_size % 2 == 0) {
		kernel_size -= 1;
	}
	float SIGMA1 = 15;
	float SIGMA2 = 80;
	float SIGMA3 = 250;
	float q = sqrt(2.0);
	Mat F(row, col, CV_32FC1);
	Mat F1, F2, F3;
	GaussianBlur(V, F1, Size(kernel_size, kernel_size), SIGMA1 / q);
	GaussianBlur(V, F2, Size(kernel_size, kernel_size), SIGMA2 / q);
	GaussianBlur(V, F3, Size(kernel_size, kernel_size), SIGMA3 / q);
	for (int i = 0; i < row; i++) {
		for (int j = 0; j < col; j++) {
			F.at <float>(i, j) = (F1.at<float>(i, j) + F2.at<float>(i, j) + F3.at<float>(i, j)) / 3.0;
		}
	}
	float average = mean(F)[0];
	Mat out(row, col, CV_32FC1);
	for (int i = 0; i < row; i++) {
		for (int j = 0; j < col; j++) {
			float gamma = powf(0.5, (average - F.at<float>(i, j)) / average);
			out.at<float>(i, j) = powf(V.at<float>(i, j), gamma);
		}
	}
	vector <Mat> v;
	v.push_back(H);
	v.push_back(S);
	v.push_back(out);
	Mat merge_;
	merge(v, merge_);
	Mat dst = HSV2RGB(merge_);
	return dst;
}

python实现

import cv2
import numpy as np


def Adaptive_light_correction(img):
    height = img.shape[0]
    width = img.shape[1]

    HSV_img = cv2.cvtColor(img,cv2.COLOR_BGR2HSV)
    V = HSV_img[:,:,2]

    kernel_size = min(height, width)
    if kernel_size % 2 == 0:
        kernel_size -= 1

    SIGMA1 = 15
    SIGMA2 = 80
    SIGMA3 = 250
    q = np.sqrt(2.0)
    F = np.zeros((height,width,3),dtype=np.float64)
    F[:,:,0] = cv2.GaussianBlur(V,(kernel_size, kernel_size),SIGMA1 / q)
    F[:,:,1] = cv2.GaussianBlur(V,(kernel_size, kernel_size),SIGMA2 / q)
    F[:,:,2] = cv2.GaussianBlur(V,(kernel_size, kernel_size),SIGMA3 / q)
    F_mean = np.mean(F,axis=2)
    average = np.mean(F_mean)
    gamma = np.power(0.5,np.divide(np.subtract(average,F_mean),average))
    out = np.power(V/255.0,gamma)*255.0
    HSV_img[:,:,2] = out
    img = cv2.cvtColor(HSV_img,cv2.COLOR_HSV2BGR)
    return img


if __name__ == '__main__':
    img = cv2.imread("./112.jpg")
    result_img = Adaptive_light_correction(img)
    cv2.imshow("result_img",result_img)
    cv2.waitKey(0)
    cv2.destroyAllWindows()

测试原图与效果如下：


参考：https://cloud.tencent.com/developer/article/1588007
特别鸣谢：https://blog.csdn.net/codeswarrior