Given an NxN binary matrix (containing only 0's or 1's), how can we go about finding largest rectangle containing all 0's?
Example:
I
0 0 0 0 1 0
0 0 1 0 0 1
II->0 0 0 0 0 0
1 0 0 0 0 0
0 0 0 0 0 1 <--IV
0 0 1 0 0 0
IV
For the above example, it is a 6×6 binary matrix. the return value in this case will be Cell 1:(2, 1) and Cell 2:(4, 4). The resulting sub-matrix can be square or rectangular. The return value can also be the size of the largest sub-matrix of all 0's, in this example 3 × 4.
Here's a solution based on the "Largest Rectangle in a Histogram" problem suggested by #j_random_hacker in the comments:
[Algorithm] works by iterating through
rows from top to bottom, for each row
solving this problem, where the
"bars" in the "histogram" consist of
all unbroken upward trails of zeros
that start at the current row (a
column has height 0 if it has a 1 in
the current row).
The input matrix mat may be an arbitrary iterable e.g., a file or a network stream. Only one row is required to be available at a time.
#!/usr/bin/env python
from collections import namedtuple
from operator import mul
Info = namedtuple('Info', 'start height')
def max_size(mat, value=0):
"""Find height, width of the largest rectangle containing all `value`'s."""
it = iter(mat)
hist = [(el==value) for el in next(it, [])]
max_size = max_rectangle_size(hist)
for row in it:
hist = [(1+h) if el == value else 0 for h, el in zip(hist, row)]
max_size = max(max_size, max_rectangle_size(hist), key=area)
return max_size
def max_rectangle_size(histogram):
"""Find height, width of the largest rectangle that fits entirely under
the histogram.
"""
stack = []
top = lambda: stack[-1]
max_size = (0, 0) # height, width of the largest rectangle
pos = 0 # current position in the histogram
for pos, height in enumerate(histogram):
start = pos # position where rectangle starts
while True:
if not stack or height > top().height:
stack.append(Info(start, height)) # push
elif stack and height < top().height:
max_size = max(max_size, (top().height, (pos - top().start)),
key=area)
start, _ = stack.pop()
continue
break # height == top().height goes here
pos += 1
for start, height in stack:
max_size = max(max_size, (height, (pos - start)), key=area)
return max_size
def area(size):
return reduce(mul, size)
The solution is O(N), where N is the number of elements in a matrix. It requires O(ncols) additional memory, where ncols is the number of columns in a matrix.
Latest version with tests is at https://gist.github.com/776423
Please take a look at Maximize the rectangular area under Histogram and then continue reading the solution below.
Traverse the matrix once and store the following;
For x=1 to N and y=1 to N
F[x][y] = 1 + F[x][y-1] if A[x][y] is 0 , else 0
Then for each row for x=N to 1
We have F[x] -> array with heights of the histograms with base at x.
Use O(N) algorithm to find the largest area of rectangle in this histogram = H[x]
From all areas computed, report the largest.
Time complexity is O(N*N) = O(N²) (for an NxN binary matrix)
Example:
Initial array F[x][y] array
0 0 0 0 1 0 1 1 1 1 0 1
0 0 1 0 0 1 2 2 0 2 1 0
0 0 0 0 0 0 3 3 1 3 2 1
1 0 0 0 0 0 0 4 2 4 3 2
0 0 0 0 0 1 1 5 3 5 4 0
0 0 1 0 0 0 2 6 0 6 5 1
For x = N to 1
H[6] = 2 6 0 6 5 1 -> 10 (5*2)
H[5] = 1 5 3 5 4 0 -> 12 (3*4)
H[4] = 0 4 2 4 3 2 -> 10 (2*5)
H[3] = 3 3 1 3 2 1 -> 6 (3*2)
H[2] = 2 2 0 2 1 0 -> 4 (2*2)
H[1] = 1 1 1 1 0 1 -> 4 (1*4)
The largest area is thus H[5] = 12
Here is a Python3 solution, which returns the position in addition to the area of the largest rectangle:
#!/usr/bin/env python3
import numpy
s = '''0 0 0 0 1 0
0 0 1 0 0 1
0 0 0 0 0 0
1 0 0 0 0 0
0 0 0 0 0 1
0 0 1 0 0 0'''
nrows = 6
ncols = 6
skip = 1
area_max = (0, [])
a = numpy.fromstring(s, dtype=int, sep=' ').reshape(nrows, ncols)
w = numpy.zeros(dtype=int, shape=a.shape)
h = numpy.zeros(dtype=int, shape=a.shape)
for r in range(nrows):
for c in range(ncols):
if a[r][c] == skip:
continue
if r == 0:
h[r][c] = 1
else:
h[r][c] = h[r-1][c]+1
if c == 0:
w[r][c] = 1
else:
w[r][c] = w[r][c-1]+1
minw = w[r][c]
for dh in range(h[r][c]):
minw = min(minw, w[r-dh][c])
area = (dh+1)*minw
if area > area_max[0]:
area_max = (area, [(r-dh, c-minw+1, r, c)])
print('area', area_max[0])
for t in area_max[1]:
print('Cell 1:({}, {}) and Cell 2:({}, {})'.format(*t))
Output:
area 12
Cell 1:(2, 1) and Cell 2:(4, 4)
Here is J.F. Sebastians method translated into C#:
private Vector2 MaxRectSize(int[] histogram) {
Vector2 maxSize = Vector2.zero;
int maxArea = 0;
Stack<Vector2> stack = new Stack<Vector2>();
int x = 0;
for (x = 0; x < histogram.Length; x++) {
int start = x;
int height = histogram[x];
while (true) {
if (stack.Count == 0 || height > stack.Peek().y) {
stack.Push(new Vector2(start, height));
} else if(height < stack.Peek().y) {
int tempArea = (int)(stack.Peek().y * (x - stack.Peek().x));
if(tempArea > maxArea) {
maxSize = new Vector2(stack.Peek().y, (x - stack.Peek().x));
maxArea = tempArea;
}
Vector2 popped = stack.Pop();
start = (int)popped.x;
continue;
}
break;
}
}
foreach (Vector2 data in stack) {
int tempArea = (int)(data.y * (x - data.x));
if(tempArea > maxArea) {
maxSize = new Vector2(data.y, (x - data.x));
maxArea = tempArea;
}
}
return maxSize;
}
public Vector2 GetMaximumFreeSpace() {
// STEP 1:
// build a seed histogram using the first row of grid points
// example: [true, true, false, true] = [1,1,0,1]
int[] hist = new int[gridSizeY];
for (int y = 0; y < gridSizeY; y++) {
if(!invalidPoints[0, y]) {
hist[y] = 1;
}
}
// STEP 2:
// get a starting max area from the seed histogram we created above.
// using the example from above, this value would be [1, 1], as the only valid area is a single point.
// another example for [0,0,0,1,0,0] would be [1, 3], because the largest area of contiguous free space is 3.
// Note that at this step, the heigh fo the found rectangle will always be 1 because we are operating on
// a single row of data.
Vector2 maxSize = MaxRectSize(hist);
int maxArea = (int)(maxSize.x * maxSize.y);
// STEP 3:
// build histograms for each additional row, re-testing for new possible max rectangluar areas
for (int x = 1; x < gridSizeX; x++) {
// build a new histogram for this row. the values of this row are
// 0 if the current grid point is occupied; otherwise, it is 1 + the value
// of the previously found historgram value for the previous position.
// What this does is effectly keep track of the height of continous avilable spaces.
// EXAMPLE:
// Given the following grid data (where 1 means occupied, and 0 means free; for clairty):
// INPUT: OUTPUT:
// 1.) [0,0,1,0] = [1,1,0,1]
// 2.) [0,0,1,0] = [2,2,0,2]
// 3.) [1,1,0,1] = [0,0,1,0]
//
// As such, you'll notice position 1,0 (row 1, column 0) is 2, because this is the height of contiguous
// free space.
for (int y = 0; y < gridSizeY; y++) {
if(!invalidPoints[x, y]) {
hist[y] = 1 + hist[y];
} else {
hist[y] = 0;
}
}
// find the maximum size of the current histogram. If it happens to be larger
// that the currently recorded max size, then it is the new max size.
Vector2 maxSizeTemp = MaxRectSize(hist);
int tempArea = (int)(maxSizeTemp.x * maxSizeTemp.y);
if (tempArea > maxArea) {
maxSize = maxSizeTemp;
maxArea = tempArea;
}
}
// at this point, we know the max size
return maxSize;
}
A few things to note about this:
This version is meant for use with the Unity API. You can easily make this more generic by replacing instances of Vector2 with KeyValuePair. Vector2 is only used for a convenient way to store two values.
invalidPoints[] is an array of bool, where true means the grid point is "in use", and false means it is not.
Solution with space complexity O(columns) [Can be modified to O(rows) also] and time complexity O(rows*columns)
public int maximalRectangle(char[][] matrix) {
int m = matrix.length;
if (m == 0)
return 0;
int n = matrix[0].length;
int maxArea = 0;
int[] aux = new int[n];
for (int i = 0; i < n; i++) {
aux[i] = 0;
}
for (int i = 0; i < m; i++) {
for (int j = 0; j < n; j++) {
aux[j] = matrix[i][j] - '0' + aux[j];
maxArea = Math.max(maxArea, maxAreaHist(aux));
}
}
return maxArea;
}
public int maxAreaHist(int[] heights) {
int n = heights.length;
Stack<Integer> stack = new Stack<Integer>();
stack.push(0);
int maxRect = heights[0];
int top = 0;
int leftSideArea = 0;
int rightSideArea = heights[0];
for (int i = 1; i < n; i++) {
if (stack.isEmpty() || heights[i] >= heights[stack.peek()]) {
stack.push(i);
} else {
while (!stack.isEmpty() && heights[stack.peek()] > heights[i]) {
top = stack.pop();
rightSideArea = heights[top] * (i - top);
leftSideArea = 0;
if (!stack.isEmpty()) {
leftSideArea = heights[top] * (top - stack.peek() - 1);
} else {
leftSideArea = heights[top] * top;
}
maxRect = Math.max(maxRect, leftSideArea + rightSideArea);
}
stack.push(i);
}
}
while (!stack.isEmpty()) {
top = stack.pop();
rightSideArea = heights[top] * (n - top);
leftSideArea = 0;
if (!stack.isEmpty()) {
leftSideArea = heights[top] * (top - stack.peek() - 1);
} else {
leftSideArea = heights[top] * top;
}
maxRect = Math.max(maxRect, leftSideArea + rightSideArea);
}
return maxRect;
}
But I get Time Limite exceeded excpetion when I try this on LeetCode. Is there any less complex solution?
I propose a O(nxn) method.
First, you can list all the maximum empty rectangles. Empty means that it covers only 0s. A maximum empty rectangle is such that it cannot be extended in a direction without covering (at least) one 1.
A paper presenting a O(nxn) algorithm to create such a list can be found at www.ulg.ac.be/telecom/rectangles as well as source code (not optimized). There is no need to store the list, it is sufficient to call a callback function each time a rectangle is found by the algorithm, and to store only the largest one (or choose another criterion if you want).
Note that a proof exists (see the paper) that the number of largest empty rectangles is bounded by the number of pixels of the image (nxn in this case).
Therefore, selecting the optimal rectangle can be done in O(nxn), and the overall method is also O(nxn).
In practice, this method is very fast, and is used for realtime video stream analysis.
Here is a version of jfs' solution, which also delivers the position of the largest rectangle:
from collections import namedtuple
from operator import mul
Info = namedtuple('Info', 'start height')
def max_rect(mat, value=0):
"""returns (height, width, left_column, bottom_row) of the largest rectangle
containing all `value`'s.
Example:
[[0, 0, 0, 0, 0, 0, 0, 0, 3, 2],
[0, 4, 0, 2, 4, 0, 0, 1, 0, 0],
[1, 0, 1, 0, 0, 0, 3, 0, 0, 4],
[0, 0, 0, 0, 4, 2, 0, 0, 0, 0],
[0, 0, 0, 2, 0, 0, 0, 0, 0, 0],
[4, 3, 0, 0, 1, 2, 0, 0, 0, 0],
[3, 0, 0, 0, 2, 0, 0, 0, 0, 4],
[0, 0, 0, 1, 0, 3, 2, 4, 3, 2],
[0, 3, 0, 0, 0, 2, 0, 1, 0, 0]]
gives: (3, 4, 6, 5)
"""
it = iter(mat)
hist = [(el==value) for el in next(it, [])]
max_rect = max_rectangle_size(hist) + (0,)
for irow,row in enumerate(it):
hist = [(1+h) if el == value else 0 for h, el in zip(hist, row)]
max_rect = max(max_rect, max_rectangle_size(hist) + (irow+1,), key=area)
# irow+1, because we already used one row for initializing max_rect
return max_rect
def max_rectangle_size(histogram):
stack = []
top = lambda: stack[-1]
max_size = (0, 0, 0) # height, width and start position of the largest rectangle
pos = 0 # current position in the histogram
for pos, height in enumerate(histogram):
start = pos # position where rectangle starts
while True:
if not stack or height > top().height:
stack.append(Info(start, height)) # push
elif stack and height < top().height:
max_size = max(max_size, (top().height, (pos - top().start), top().start), key=area)
start, _ = stack.pop()
continue
break # height == top().height goes here
pos += 1
for start, height in stack:
max_size = max(max_size, (height, (pos - start), start), key=area)
return max_size
def area(size):
return size[0] * size[1]
To be complete, here's the C# version which outputs the rectangle coordinates.
It's based on dmarra's answer but without any other dependencies.
There's only the function bool GetPixel(int x, int y), which returns true when a pixel is set at the coordinates x,y.
public struct INTRECT
{
public int Left, Right, Top, Bottom;
public INTRECT(int aLeft, int aTop, int aRight, int aBottom)
{
Left = aLeft;
Top = aTop;
Right = aRight;
Bottom = aBottom;
}
public int Width { get { return (Right - Left + 1); } }
public int Height { get { return (Bottom - Top + 1); } }
public bool IsEmpty { get { return Left == 0 && Right == 0 && Top == 0 && Bottom == 0; } }
public static bool operator ==(INTRECT lhs, INTRECT rhs)
{
return lhs.Left == rhs.Left && lhs.Top == rhs.Top && lhs.Right == rhs.Right && lhs.Bottom == rhs.Bottom;
}
public static bool operator !=(INTRECT lhs, INTRECT rhs)
{
return !(lhs == rhs);
}
public override bool Equals(Object obj)
{
return obj is INTRECT && this == (INTRECT)obj;
}
public bool Equals(INTRECT obj)
{
return this == obj;
}
public override int GetHashCode()
{
return Left.GetHashCode() ^ Right.GetHashCode() ^ Top.GetHashCode() ^ Bottom.GetHashCode();
}
}
public INTRECT GetMaximumFreeRectangle()
{
int XEnd = 0;
int YStart = 0;
int MaxRectTop = 0;
INTRECT MaxRect = new INTRECT();
// STEP 1:
// build a seed histogram using the first row of grid points
// example: [true, true, false, true] = [1,1,0,1]
int[] hist = new int[Height];
for (int y = 0; y < Height; y++)
{
if (!GetPixel(0, y))
{
hist[y] = 1;
}
}
// STEP 2:
// get a starting max area from the seed histogram we created above.
// using the example from above, this value would be [1, 1], as the only valid area is a single point.
// another example for [0,0,0,1,0,0] would be [1, 3], because the largest area of contiguous free space is 3.
// Note that at this step, the heigh fo the found rectangle will always be 1 because we are operating on
// a single row of data.
Tuple<int, int> maxSize = MaxRectSize(hist, out YStart);
int maxArea = (int)(maxSize.Item1 * maxSize.Item2);
MaxRectTop = YStart;
// STEP 3:
// build histograms for each additional row, re-testing for new possible max rectangluar areas
for (int x = 1; x < Width; x++)
{
// build a new histogram for this row. the values of this row are
// 0 if the current grid point is occupied; otherwise, it is 1 + the value
// of the previously found historgram value for the previous position.
// What this does is effectly keep track of the height of continous avilable spaces.
// EXAMPLE:
// Given the following grid data (where 1 means occupied, and 0 means free; for clairty):
// INPUT: OUTPUT:
// 1.) [0,0,1,0] = [1,1,0,1]
// 2.) [0,0,1,0] = [2,2,0,2]
// 3.) [1,1,0,1] = [0,0,1,0]
//
// As such, you'll notice position 1,0 (row 1, column 0) is 2, because this is the height of contiguous
// free space.
for (int y = 0; y < Height; y++)
{
if (!GetPixel(x, y))
{
hist[y]++;
}
else
{
hist[y] = 0;
}
}
// find the maximum size of the current histogram. If it happens to be larger
// that the currently recorded max size, then it is the new max size.
Tuple<int, int> maxSizeTemp = MaxRectSize(hist, out YStart);
int tempArea = (int)(maxSizeTemp.Item1 * maxSizeTemp.Item2);
if (tempArea > maxArea)
{
maxSize = maxSizeTemp;
maxArea = tempArea;
MaxRectTop = YStart;
XEnd = x;
}
}
MaxRect.Left = XEnd - maxSize.Item1 + 1;
MaxRect.Top = MaxRectTop;
MaxRect.Right = XEnd;
MaxRect.Bottom = MaxRectTop + maxSize.Item2 - 1;
// at this point, we know the max size
return MaxRect;
}
private Tuple<int, int> MaxRectSize(int[] histogram, out int YStart)
{
Tuple<int, int> maxSize = new Tuple<int, int>(0, 0);
int maxArea = 0;
Stack<Tuple<int, int>> stack = new Stack<Tuple<int, int>>();
int x = 0;
YStart = 0;
for (x = 0; x < histogram.Length; x++)
{
int start = x;
int height = histogram[x];
while (true)
{
if (stack.Count == 0 || height > stack.Peek().Item2)
{
stack.Push(new Tuple<int, int>(start, height));
}
else if (height < stack.Peek().Item2)
{
int tempArea = (int)(stack.Peek().Item2 * (x - stack.Peek().Item1));
if (tempArea > maxArea)
{
YStart = stack.Peek().Item1;
maxSize = new Tuple<int, int>(stack.Peek().Item2, (x - stack.Peek().Item1));
maxArea = tempArea;
}
Tuple<int, int> popped = stack.Pop();
start = (int)popped.Item1;
continue;
}
break;
}
}
foreach (Tuple<int, int> data in stack)
{
int tempArea = (int)(data.Item2 * (x - data.Item1));
if (tempArea > maxArea)
{
YStart = data.Item1;
maxSize = new Tuple<int, int>(data.Item2, (x - data.Item1));
maxArea = tempArea;
}
}
return maxSize;
}
An appropriate algorithm can be found within Algorithm for finding the largest inscribed rectangle in polygon (2019).
I implemented it in python:
import largestinteriorrectangle as lir
import numpy as np
grid = np.array([[0, 0, 0, 0, 1, 0],
[0, 0, 1, 0, 0, 1],
[0, 0, 0, 0, 0, 0],
[1, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 1],
[0, 0, 1, 0, 0, 0]],
"bool")
grid = ~grid
lir.lir(grid) # [1, 2, 4, 3]
the result comes as x, y, width, height
I am building a program in objective C/C++ and openCV. I am pretty skilled in Objective C but new to C++.
I am building custom RGB2HSV algorithm. My algorithm is slightly different from the openCV library cvtColor(in, out, CV_RGB2HSV).
The one I try to translate form Matlab to opencV/C++ produces so clear HSV image that no additional filtering is needed before further processing. Code below – Matlab code is self-explanatory.
I try to translate it to C++/openCV function out of it but I hit the wall trying to access pixel values of the image. I am new to C++.
I read a lot on the ways how to access Mat structure but usually I obtain either bunch of letters in a place of zeros or a number typically something like this “\202 k g”. When I try to do any multiplication operations on the say \202 the result has nothing to do with math.
Please help me to properly access the pixel values. Also in current version using uchar won’t work because some values are outside 0-255 range.
The algorithm is not mine. I cannot even point the source but it gives clearly better results than stock RGB2HSV.
Also the algorithm below is for one pixel. It needs to be applied each pixel in the image so in final version it need to wrapped with for { for {}} loops.
I also wish to share this method with community so everyone can benefit from it and saving on pre-filtering.
Please help me translate it to C++ / openCV. If possible with the best practices speed wise. Or at least how to clearly access the pixel value so it is workable with range of mathematical equations. Thanks in advance.
function[H, S, V] = rgb2hsvPixel(R,G,B)
% Algorithm:
% In case of 8-bit and 16-bit images, `R`, `G`, and `B` are converted to the
% floating-point format and scaled to fit the 0 to 1 range.
%
% V = max(R,G,B)
% S = / (V - min(R,G,B)) / V if V != 0
% \ 0 otherwise
% / 60*(G-B) / (V - min(R,G,B)) if V=R
% H = | 120 + 60*(B-R) / (V - min(R,G,B)) if V=G
% \ 240 + 60*(R-G) / (V - min(R,G,B)) if V=B
%
% If `H<0` then `H=H+360`. On output `0<=V<=1`, `0<=S<=1`, `0<=H<=360`.
red = (double(R)-16)*255/224; % \
green = (double(G)-16)*255/224; % }- R,G,B (0 <-> 255) -> (-18.2143 <-> 272.0759)
blue = (min(double(B)*2,240)-16)*255/224; % /
minV = min(red,min(green,blue));
value = max(red,max(green,blue));
delta = value - minV;
if(value~=0)
sat = (delta*255) / value;% s
if (delta ~= 0)
if( red == value )
hue = 60*( green - blue ) / delta; % between yellow & magenta
elseif( green == value )
hue = 120 + 60*( blue - red ) / delta; % between cyan & yellow
else
hue = 240 + 60*( red - green ) / delta; % between magenta & cyan
end
if( hue < 0 )
hue = hue + 360;
end
else
hue = 0;
sat = 0;
end
else
% r = g = b = 0
sat = 0;
hue = 0;
end
H = max(min(floor(((hue*255)/360)),255),0);
S = max(min(floor(sat),255),0);
V = max(min(floor(value),255),0);
end
To access the value of a pixel in a 3-channel, 8-bit precision image (type CV_8UC3) you have to do it like this:
cv::Mat image;
cv::Vec3b BGR = image.at<cv::Vec3b>(i,j);
If, as you say, 8-bit precision and range are not enough, you can declare a cv::Mat of type CV_32F to store floating point 32-bit numbers.
cv::Mat image(height, width, CV_32FC3);
//fill your image with data
for(int i = 0; i < image.rows; i++) {
for(int j = 0; j < image.cols; j++) {
cv::Vec3f BGR = image.at<cv::Vec3f>(i,j)
//process your pixel
cv::Vec3f HSV; //your calculated HSV values
image.at<cv::Vec3f>(i,j) = HSV;
}
}
Be aware that OpenCV stores rgb values in the BGR order and not RGB. Take a look at OpenCV docs to learn more about it.
If you are concerned by performance and fairly comfortable with pixel indexes, you can use directly the Mat ptr.
For example:
cv::Mat img = cv::Mat::zeros(4, 8, CV_8UC3);
uchar *ptr_row_img;
int cpt = 0;
for(int i = 0; i < img.rows; i++) {
ptr_row_img = img.ptr<uchar>(i);
for(int j = 0; j < img.cols; j++) {
for(int c = 0; c < img.channels(); c++, cpt++, ++ptr_row_img) {
*ptr_row_img = cpt;
}
}
}
std::cout << "img=\n" << img << std::endl;
The previous code should print:
img= [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12,
13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23; 24, 25, 26,
27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40,
41, 42, 43, 44, 45, 46, 47; 48, 49, 50, 51, 52, 53, 54,
55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68,
69, 70, 71; 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82,
83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95]
The at access should be enough for most of the cases and is much more readable / less likely to make a mistake than using the ptr access.
References:
How to scan images, lookup tables and time measurement with OpenCV
C++: OpenCV: fast pixel iteration
Thanks everybody for help.
Thanks to your hints I constructed the custom rgb2hsv function C++/openCV.
From the top left respectively, edges after bgr->gray->edges, bgr->HSV->edges, bgr->customHSV->edges
Below each of them corresponding settings of the filters to achieve approximately the same clear results. The bigger the radius of a filter the more complex and time consuming computations.
It produces clearer edges in next steps of image processing.
It can be tweaked further experimenting with parameters in r g b channels:
red = (red-16)*1.1384; //255/244=1.1384
here 16 – the bigger the number the clearer V becomes
255/244 – also affect the outcome extending it beyond ranges 0-255, later to be clipped.
This numbers here seem to be golden ratio but anyone can adjust for specific needs.
With this function translating BGR to RGB can be avoided by directly connecting colors to proper channels in raw image.
Probably it is a little clumsy performance wise. In my case it serves in first step of color balance and histogram adjustment so speed is not that critical.
To use in constant processing video stream it need speed optimization, I think by using pointers and reducing loop complexity. Optimization is not exactly my cup of tea. So if someone helped to optimize it for the community that would be great.
Here it is ready to use:
Mat bgr2hsvCustom ( Mat& image )
{
//smallParam = 16;
for(int x = 0; x < image.rows; x++)
{
for(int y = 0; y<image.cols; y++)
{
//assigning vector to individual float BGR values
float blue = image.at<cv::Vec3b>(x,y)[0];
float green = image.at<cv::Vec3b>(x,y)[1];
float red = image.at<cv::Vec3b>(x,y)[2];
float sat, hue, minValue, maxValue, delta;
float const ang0 = 0; // func min and max don't accept varaible and number
float const ang240 = 240;
float const ang255 = 255;
red = (red-16)*1.1384; //255/244
green = (green-16)*1.1384;
blue = (min(blue*2,ang240)-16)*1.1384;
minValue = min(red,min(green,blue));
maxValue = max(red,max(green,blue));
delta = maxValue - minValue;
if (maxValue != 0)
{
sat = (delta*255) / maxValue;
if ( delta != 0)
{
if (red == maxValue){
hue = 60*(green - blue)/delta;
}
else if( green == maxValue ) {
hue = 120 + 60*( blue - red )/delta;
}
else{
hue = 240 + 60*( red - green )/delta;
}
if( hue < 0 ){
hue = hue + 360;
}
}
else{
sat = 0;
hue = 0;
}
}
else{
hue = 0;
sat = 0;
}
image.at<cv::Vec3b>(x,y)[0] = max(min(floor(maxValue),ang255),ang0); //V
image.at<cv::Vec3b>(x,y)[1] = max(min(floor(sat),ang255),ang0); //S
image.at<cv::Vec3b>(x,y)[2] = max(min(floor(((hue*255)/360)),ang255),ang0); //H
}
}
return image;
}
I want calculate angles of gradients from depth map and group it for some directions (8 sectors)
But my function calculates only first 3 directions
cv::Mat calcAngles(cv::Mat dimg)//dimg is depth map
{
const int directions_num = 8;//number of directions
const int degree_grade = 360;
int range_coeff = 255 / (directions_num + 1);//just for visualize
cv::Mat x_edge, y_edge, full_edge, angles;
dimg.copyTo(x_edge);
dimg.copyTo(y_edge);
dimg.copyTo(full_edge);
//compute gradients
Sobel( dimg, x_edge, CV_8U, 1, 0, 5, 1, 19, 4 );
Sobel( dimg, y_edge, CV_8U, 0, 1, 5, 1, 19, 4 );
Sobel( dimg, full_edge, CV_8U, 1, 1, 5, 1, 19, 4 );
float freq[directions_num + 1];//for collect direction's frequency
memset(freq, 0, sizeof(freq));
angles = cv::Mat::zeros(dimg.rows, dimg.cols, CV_8U);//store directions here
for(int i = 0; i < angles.rows; i++)
{
for(int j = 0; j < angles.cols; j++)
{
angles.at<uchar>(i, j) = (((int)cv::fastAtan2(y_edge.at<uchar>(i, j), x_edge.at<uchar>(i, j))) / (degree_grade/directions_num) + 1
) * (dimg.at<uchar>(i, j) ? 1 : 0);//fastatan returns values from 0 to 360, if i not mistaken. I want group angles by directions_num sectors. I use first 'direction' (zero value) for zero values from depth map (zero value at my depth map suggest that it is bad pixel)
freq[angles.at<uchar>(i, j)] += 1;
}
}
for(int i = 0; i < directions_num + 1; i++)
{
printf("%2.2f\t", freq[i]);
}
printf("\n");
angles *= range_coeff;//for visualization
return angles;
}
Out from one of the frames:
47359.00 15018.00 8199.00 6224.00 0.00 0.00 0.00 0.00 0.00
(first value is "zero pixel", next is number of gradients in n-place but only 3 are not zero)
Visualization
Is there way out? Or these result is OK?
PS Sorry for my writing mistakes. English in not my native language.
You used CV_8U type for Sobel output. It is unsigned integer 8 bit. So it can store only positive values. That's why fastAtan2 returns less or equal than 90. Change type to CV_16S and use short type for accessing the elements:
cv::Sobel(dimg, x_edge, CV_16S, 1, 0, 5, 1, 19, 4);
cv::Sobel(dimg, y_edge, CV_16S, 0, 1, 5, 1, 19, 4);
cv::fastAtan2(y_edge.at<short>(i, j), x_edge.at<short>(i, j))
At Wikipedia's Mandelbrot set page there are really beautiful generated images of the Mandelbrot set.
I also just implemented my own Mandelbrot algorithm. Given n is the number of iterations used to calculate each pixel, I color them pretty simple from black to green to white like that (with C++ and Qt 5.0):
QColor mapping(Qt::white);
if (n <= MAX_ITERATIONS){
double quotient = (double) n / (double) MAX_ITERATIONS;
double color = _clamp(0.f, 1.f, quotient);
if (quotient > 0.5) {
// Close to the mandelbrot set the color changes from green to white
mapping.setRgbF(color, 1.f, color);
}
else {
// Far away it changes from black to green
mapping.setRgbF(0.f, color, 0.f);
}
}
return mapping;
My result looks like that:
I like it pretty much already, but which color gradient is used for the images in Wikipedia? How to calculate that gradient with a given n of iterations?
(This question is not about smoothing.)
The gradient is probably from Ultra Fractal. It is defined by 5 control points:
Position = 0.0 Color = ( 0, 7, 100)
Position = 0.16 Color = ( 32, 107, 203)
Position = 0.42 Color = (237, 255, 255)
Position = 0.6425 Color = (255, 170, 0)
Position = 0.8575 Color = ( 0, 2, 0)
where Position is in range [0, 1) and Color is RGB in range [0, 255].
The catch is that the colors are not linearly interpolated. The interpolation of colors is likely cubic (or something similar). Following image shows the difference between linear and Monotone cubic interpolation:
As you can see the cubic interpolation results in smoother and "prettier" gradient. I used monotone cubic interpolation to avoid "overshooting" of the [0, 255] color range that can be caused by cubic interpolation. Monotone cubic ensures that interpolated values are always in the range of input points.
I use following code to compute the color based on iteration i:
double smoothed = Math.Log2(Math.Log2(re * re + im * im) / 2); // log_2(log_2(|p|))
int colorI = (int)(Math.Sqrt(i + 10 - smoothed) * gradient.Scale) % colors.Length;
Color color = colors[colorI];
where i is the diverged iteration number, re and im are diverged coordinates, gradient.Scale is 256, and the colors is and array with pre-computed gradient colors showed above. Its length is 2048 in this case.
Well, I did some reverse engineering on the colours used in wikipedia using the Photoshop eyedropper. There are 16 colours in this gradient:
R G B
66 30 15 # brown 3
25 7 26 # dark violett
9 1 47 # darkest blue
4 4 73 # blue 5
0 7 100 # blue 4
12 44 138 # blue 3
24 82 177 # blue 2
57 125 209 # blue 1
134 181 229 # blue 0
211 236 248 # lightest blue
241 233 191 # lightest yellow
248 201 95 # light yellow
255 170 0 # dirty yellow
204 128 0 # brown 0
153 87 0 # brown 1
106 52 3 # brown 2
Simply using a modulo and an QColor array allows me to iterate through all colours in the gradient:
if (n < MAX_ITERATIONS && n > 0) {
int i = n % 16;
QColor mapping[16];
mapping[0].setRgb(66, 30, 15);
mapping[1].setRgb(25, 7, 26);
mapping[2].setRgb(9, 1, 47);
mapping[3].setRgb(4, 4, 73);
mapping[4].setRgb(0, 7, 100);
mapping[5].setRgb(12, 44, 138);
mapping[6].setRgb(24, 82, 177);
mapping[7].setRgb(57, 125, 209);
mapping[8].setRgb(134, 181, 229);
mapping[9].setRgb(211, 236, 248);
mapping[10].setRgb(241, 233, 191);
mapping[11].setRgb(248, 201, 95);
mapping[12].setRgb(255, 170, 0);
mapping[13].setRgb(204, 128, 0);
mapping[14].setRgb(153, 87, 0);
mapping[15].setRgb(106, 52, 3);
return mapping[i];
}
else return Qt::black;
The result looks pretty much like what I was looking for:
:)
I believe they're the default colours in Ultra Fractal. The evaluation version comes with source for a lot of the parameters, and I think that includes that colour map (if you can't infer it from the screenshot on the front page) and possibly also the logic behind dynamically scaling that colour map appropriately for each scene.
This is an extension of NightElfik's great answer.
The python library Scipy has monotone cubic interpolation methods in version 1.5.2 with pchip_interpolate. I included the code I used to create my gradient below. I decided to include helper values less than 0 and larger than 1 to help the interpolation wrap from the end to the beginning (no sharp corners).
#set up the control points for your gradient
yR_observed = [0, 0,32,237, 255, 0, 0, 32]
yG_observed = [2, 7, 107, 255, 170, 2, 7, 107]
yB_observed = [0, 100, 203, 255, 0, 0, 100, 203]
x_observed = [-.1425, 0, .16, .42, .6425, .8575, 1, 1.16]
#Create the arrays with the interpolated values
x = np.linspace(min(x_observed), max(x_observed), num=1000)
yR = pchip_interpolate(x_observed, yR_observed, x)
yG = pchip_interpolate(x_observed, yG_observed, x)
yB = pchip_interpolate(x_observed, yB_observed, x)
#Convert them back to python lists
x = list(x)
yR = list(yR)
yG = list(yG)
yB = list(yB)
#Find the indexs where x crosses 0 and crosses 1 for slicing
start = 0
end = 0
for i in x:
if i > 0:
start = x.index(i)
break
for i in x:
if i > 1:
end = x.index(i)
break
#Slice away the helper data in the begining and end leaving just 0 to 1
x = x[start:end]
yR = yR[start:end]
yG = yG[start:end]
yB = yB[start:end]
#Plot the values if you want
#plt.plot(x, yR, color = "red")
#plt.plot(x, yG, color = "green")
#plt.plot(x, yB, color = "blue")
#plt.show()