OpenCV：对16位灰度图像进行归一化处理会产生较弱的结果

Question

I want to stretch contrast in 16 bit grayscale image. but void normalize(InputArray src, OutputArray dst, double alpha=1, double beta=0, int norm_type=NORM_L2, int dtype=-1, InputArray mask=noArray() ) gives me little bit brighter image, but still too dark.

Documentation: http://docs.opencv.org/2.4/modules/core/doc/operations_on_arrays.html#normalize

says that alpha is the lower limit and beta upper limit. So in case of 16 bit image i would expect 0 and 65535.0 are correct values. I did a research and most answers pointed out that alpha and betha are minimum and maximum in normalized image.

#include "stdafx.h"
#include <iostream>
#include <opencv2/opencv.hpp>
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
int main()
 {
    cv::Mat image;
    image = cv::imread("darkImage.tif", CV_LOAD_IMAGE_ANYDEPTH | CV_LOAD_IMAGE_GRAYSCALE);

if (!image.data)                           
    {
        std::cout << "Could not open or find the image" << std::endl;
        return -1;
    }

    cv::namedWindow("Original", CV_WINDOW_NORMAL | CV_WINDOW_KEEPRATIO);
    cv::imshow("Original", image);

    cv::normalize(image, image, 0, 65535.0, cv::NORM_MINMAX, CV_16U);
    cv::namedWindow("Normalize", CV_WINDOW_NORMAL | CV_WINDOW_KEEPRATIO);
    cv::imshow("Normalize", image);
    cv::waitKey();

    return 0;

}

Original and Normalized image show that contrast enhancement is not sufficient. ImageJ normalization gives me much better result .

Are alpha and beta values appropriate for 16bit image? I am new in opencv and any help is appresiated.

I use: opencv3.1, VisualStudio2015, W10, 64bit

Answer 1

Yeah, probably histogram equaliztion is the way to go. EqualizeHist doesn't work for 16-bit. So I would recommend either

image.convertTo(image,CV_8U,1./256.);

or

image.convertTo(image,CV_32F);

followed by

equalizeHist(image,imageEq);

The 8-bit option is tried and true, but might lose information during truncation. I haven't tried this for float, but I suspect it ends up truncating/binning internally anyway, which defeats the purpose of float.

Alternatively, if you wanted to do it properly and weren't concerned with run-time/dev-time, you could implement a 16-bit histogram followed by a 16bit-to-8bit lookup table, following the idea behind histogram equalization/CLAHE. (Make the cumulative distribution function and apply this 65K vector directly as a lookup table to the image to make the result uniformly distributed.)

Or if you want to do your own version of cv::normalize() to give centered mean and reasonable standard deviation, you could do something like this:

Scalar imMean, imStd;
meanStdDev(image, imMean, imStd);
double a = (1<<16)*(0.25/imStd.val[0]);   // give equalized image a stdDev of 0.25
double b = (1<<16)*0.5 - a*imMean.val[0]; // give equalized image a mean of 0.5
imageEq = a*image+b;