OpenCV study notes: how to scan images using a lookup table and timing

Original link: http://www.cnblogs.com/james1207/p/3266645.html

 purpose

We will explore answers to the following questions:

• How to traverse each pixel in the image?
• OpenCV matrix values ​​is how to store?
• How to test the performance of our algorithm is implemented?
• What lookup table? Why use it?

 Test Case

Here we tested, it is a simple color reduction method. If the matrix element is a single channel stored in the pixel, C or C ++ unsigned character type, then the pixel can have 256 different values. But if the three-channel image, the number of colors on this storage format of too much (or rather, there are more than sixteen million kinds). We could have a serious impact on the performance of the algorithm used so many colors. In fact, sometimes, only a small fraction of these colors is enough to achieve the same effect.

In this case, a method is commonly used  color space reduction  . The approach is: the existing color space is divided by an input value to obtain a small number of colors. For example, the color values of 0 to 9, preferably 0,10 to 19 the new value is preferably 10, and so on.

UCHAR  (unsigned char, i.e. the values between 0 and 255 number) divided by the value of type  int  value, the result is still  char  . Char type as a result, so also seek out a decimal is rounded down. With this, mentioned in  uchar  colors of custom fields can be expressed as reduced operation the following form:

In this case, a simple color space reduction algorithm can be represented by the following two steps: a traversing each image matrix in a pixel; Second, the application of the above formula pixel. It is noteworthy that, here we use the division and multiplication, both of which operation has been particularly time-consuming, so we should replace them with lower costs as addition, subtraction, assignment and other operations. Furthermore, it should be noted that the above-described input operation only over a limited range of values, such as  uchar  type preferably 256 values.

It can be seen, for larger images, it is effective to pre-calculate all possible values, then these values when required, using a lookup table can be assigned directly. Is a one-dimensional lookup table or multi-dimensional arrays, the output values are stored corresponding to different input values, the advantage that only read, without calculation.

Our test program (as well as sample code given here) to do the following things:

A read image (may be a color image may be a gray scale image, determined by the command line parameters), then the command line parameter is given is an integer of color reduction command-line arguments.

Currently, OpenCV has mainly three methods for traversing the image pixel by pixel. We will use these three methods were scanned image, and outputs them with the time on the screen. I think this should be very interesting contrast.

You can get from  here to  download the source code, you can also find OpenCV samples directory, enter tutorial_code of cpp's core directory, look for the code of the program. The basic usage of the program are:

how_to_scan_images imageName.jpg intValueToReduce [G]

That last argument is optional. If provided, the image loaded in grayscale format, otherwise the color format. In this program, we first calculate the lookup table.

    int divideWith; // convert our input string to number - C++ style
    stringstream s;//数据类型转换
    s << argv[2];
    s >> divideWith;
    if (!s)
    {
        cout << "Invalid number entered for dividing. " << endl; 
        return -1;
    }
    
    uchar table[256]; 
    for (int i = 0; i < 256; ++i)
       table[i] = divideWith* (i/divideWith);

这里我们先使用C++的 stringstream 类，把第三个命令行参数由字符串转换为整数。

然后，我们用数组和前面给出的公式计算查找表。这里并未涉及有关OpenCV的内容。

另外有个问题是如何计时。没错，OpenCV提供了两个简便的可用于计时的函数 getTickCount() 和 getTickFrequency() 。第一个函数返回你的CPU自某个事件（如启动电脑）以来走过的时钟周期数，第二个函数返回你的CPU一秒钟所走的时钟周期数。这样，我们就能轻松地以秒为单位对某运算计时：

double t = (double)getTickCount();
// 做点什么 ...
t = ((double)getTickCount() - t)/getTickFrequency();
cout << "Times passed in seconds: " << t << endl;

 图像矩阵是如何存储在内存之中的？

图像矩阵的大小取决于我们所用的颜色模型，确切地说，取决于所用通道数。如果是灰度图像，矩阵就会像这样：

而对多通道图像来说，矩阵中的列会包含多个子列，其子列个数与通道数相等。例如，RGB颜色模型的矩阵：

注意到，子列的通道顺序是反过来的：BGR而不是RGB。很多情况下，因为内存足够大，可实现连续存储，因此，图像中的各行就能一行一行地连接起来，形成一个长行。连续存储有助于提升图像扫描速度，我们可以使用 isContinuous() 来去判断矩阵是否是连续存储的. 相关示例会在接下来的内容中提供。

 1.高效的方法 Efficient Way

说到性能，经典的C风格运算符[]（指针）访问要更胜一筹. 因此，我们推荐的效率最高的查找表赋值方法，还是下面的这种：

Mat& ScanImageAndReduceC(Mat& I, const uchar* const table)
{
    // accept only char type matrices
    CV_Assert(I.depth() != sizeof(uchar));     

    int channels = I.channels();

    int nRows = I.rows * channels; 
    int nCols = I.cols;

    if (I.isContinuous())
    {
        nCols *= nRows;
        nRows = 1;         
    }

    int i,j;
    uchar* p; 
    for( i = 0; i < nRows; ++i)
    {
        p = I.ptr<uchar>(i);
        for ( j = 0; j < nCols; ++j)
        {
            p[j] = table[p[j]];             
        }
    }
    return I; 
}

这里，我们获取了每一行开始处的指针，然后遍历至该行末尾。如果矩阵是以连续方式存储的，我们只需请求一次指针、然后一路遍历下去就行。彩色图像的情况有必要加以注意：因为三个通道的原因，我们需要遍历的元素数目也是3倍。

这里有另外一种方法来实现遍历功能，就是使用 data ， data会从 Mat 中返回指向矩阵第一行第一列的指针。注意如果该指针为NULL则表明对象里面无输入，所以这是一种简单的检查图像是否被成功读入的方法。当矩阵是连续存储时，我们就可以通过遍历 data 来扫描整个图像。例如，一个灰度图像，其操作如下：

uchar* p = I.data;

for( unsigned int i =0; i < ncol*nrows; ++i)
    *p++ = table[*p];

这回得出和前面相同的结果。但是这种方法编写的代码可读性方面差，并且进一步操作困难。同时，我发现在实际应用中，该方法的性能表现上并不明显优于前一种（因为现在大多数编译器都会对这类操作做出优化）。

 2.迭代法 The iterator (safe) method

在高性能法（the efficient way）中，我们可以通过遍历正确的uchar域并跳过行与行之间可能的空缺-你必须自己来确认是否有空缺，来实现图像扫描，迭代法则被认为是一种以更安全的方式来实现这一功能。在迭代法中，你所需要做的仅仅是获得图像矩阵的begin和end，然后增加迭代直至从begin到end。将*操作符添加在迭代指针前，即可访问当前指向的内容。

Mat& ScanImageAndReduceIterator(Mat& I, const uchar* const table)
{
    // accept only char type matrices
    CV_Assert(I.depth() != sizeof(uchar));     
    
    const int channels = I.channels();
    switch(channels)
    {
    case 1: 
        {
            MatIterator_<uchar> it, end; 
            for( it = I.begin<uchar>(), end = I.end<uchar>(); it != end; ++it)
                *it = table[*it];
            break;
        }
    case 3: 
        {
            MatIterator_<Vec3b> it, end; 
            for( it = I.begin<Vec3b>(), end = I.end<Vec3b>(); it != end; ++it)
            {
                (*it)[0] = table[(*it)[0]];
                (*it)[1] = table[(*it)[1]];
                (*it)[2] = table[(*it)[2]];
            }
        }
    }
    
    return I; 
}

对于彩色图像中的一行，每列中有3个uchar元素，这可以被认为是一个小的包含uchar元素的vector，在OpenCV中用 Vec3b 来命名。如果要访问第n个子列，我们只需要简单的利用[]来操作就可以。需要指出的是，OpenCV的迭代在扫描过一行中所有列后会自动跳至下一行，所以说如果在彩色图像中如果只使用一个简单的 uchar 而不是 Vec3b 迭代的话就只能获得蓝色通道(B)里的值。

 3. 通过相关返回值的On-the-fly地址计算

事实上这个方法并不推荐被用来进行图像扫描，它本来是被用于获取或更改图像中的随机元素。它的基本用途是要确定你试图访问的元素的所在行数与列数。在前面的扫描方法中，我们观察到知道所查询的图像数据类型是很重要的。这里同样的你得手动指定好你要查找的数据类型。下面的代码中是一个关于灰度图像的示例(运用 + at() 函数):

Mat& ScanImageAndReduceRandomAccess(Mat& I, const uchar* const table)
{
    // accept only char type matrices
    CV_Assert(I.depth() != sizeof(uchar));     

    const int channels = I.channels();
    switch(channels)
    {
    case 1: 
        {
            for( int i = 0; i < I.rows; ++i)
                for( int j = 0; j < I.cols; ++j )
                    I.at<uchar>(i,j) = table[I.at<uchar>(i,j)];
            break;
        }
    case 3: 
        {
         Mat_<Vec3b> _I = I;
            
         for( int i = 0; i < I.rows; ++i)
            for( int j = 0; j < I.cols; ++j )
               {
                   _I(i,j)[0] = table[_I(i,j)[0]];
                   _I(i,j)[1] = table[_I(i,j)[1]];
                   _I(i,j)[2] = table[_I(i,j)[2]];
            }
         I = _I;
         break;
        }
    }
    
    return I;
}

该函数输入为数据类型及需求元素的坐标，返回的是一个对应的值-如果用 get 则是constant，如果是用 set 、则为non-constant. 出于程序安全，当且仅当在 debug 模式下 它会检查你的输入坐标是否有效或者超出范围. 如果坐标有误，则会输出一个标准的错误信息. 和高性能法（the efficient way）相比, 在 release模式下，它们之间的区别仅仅是On-the-fly方法对于图像矩阵的每个元素，都会获取一个新的行指针，通过该指针和[]操作来获取列元素。

当你对一张图片进行多次查询操作时，为避免反复输入数据类型和at带来的麻烦和浪费的时间，OpenCV 提供了:basicstructures:Mat_ <id3> data type. 它同样可以被用于获知矩阵的数据类型，你可以简单利用()操作返回值来快速获取查询结果. 值得注意的是你可以利用 at() 函数来用同样速度完成相同操作. 它仅仅是为了让懒惰的程序员少写点 >_< .

 4. 核心函数LUT（The Core Function）

这是最被推荐的用于实现批量图像元素查找和更改操作的图像方法。在图像处理中，对于一个给定的值，将其替换成其他的值是一个很常见的操作，OpenCV提供里一个函数直接实现该操作，并不需要你自己扫描图像，就是:operationsOnArrays:LUT() <lut> ，一个包含于core module的函数. 首先我们建立一个mat型用于查表:

    Mat lookUpTable(1, 256, CV_8U);
    uchar* p = lookUpTable.data; 
    for( int i = 0; i < 256; ++i)
        p[i] = table[i];

然后我们调用函数 (I 是输入 J 是输出):

        LUT(I, lookUpTable, J);

 性能表现

为了得到最优的结果，你最好自己编译并运行这些程序. 为了更好的表现性能差异，我用了一个相当大的图片(2560 X 1600). 性能测试这里用的是彩色图片，结果是数百次测试的平均值.

Efficient Way	79.4717 milliseconds
Iterator	83.7201 milliseconds
On-The-Fly RA	93.7878 milliseconds
LUT function	32.5759 milliseconds

We draw some conclusions: to make use of the built-in function called OpenCV LUT function can get the fastest since OpenCV library can enable multi-threaded architecture by Intel Thread Of course, if you prefer the method of using the pointer to the scanned image, iteration... France is a good choice, but a slower speed. Used in debug mode on-the-fly method scans the whole map is a method most waste of resources, in the release mode, it is almost the same performance and iterative methods, but from a security point of view, iterative method is a better choice

Reproduced in: https: //www.cnblogs.com/james1207/p/3266645.html