I'm extracting the HOG features of a grayscale image using OpenCV's HOG implementation. Assuming that my image matches the default window size, i.e. 128x64, I'm struggling to understand correctly how that feature vector is organised. This is what I know:
Every cell outputs a 9 elements histogram quantifying the orientations of the edges lying within that cell (8x8 pixels by default).
Each block contains 2x2 cells.
By default, an 8x8 block stride is used.
This results in a 7*15*9*4 = 3780 elements feature vector. 7 and 15 are the number of blocks that fit horizontally and vertically when a 50% block overlap is used. All great until here.
If we examine the features of the first block, i.e. the first 9*4 elements, how are they arranged? Do the first 9 bins correspond to the top left cell in the block? what about the next 9? and the next?
And which orientation angle does each of the 9 bins represent? Does bins[0] = 0, bins[1] = 20, bins[2] = 40, ... bins[8] = 160. Or is the order different, for instance going from -pi/2 to +pi/2?
Related
I have an array of data from a grayscale image that I have segmented sets of contiguous points of a certain intensity value from.
Currently I am doing a naive bounding box routine where I find the minimum and maximum (x,y) [row, col] points. This obviously does not provide the smallest possible box that contains the set of points which is demonstrable by simply rotating a rectangle so the longest axis is no longer aligned with a principal axis.
What I wish to do is find the minimum sized oriented bounding box. This seems to be possible using an algorithm known as rotating calipers, however the implementations of this algorithm seem to rely on the idea that you have a set of vertices to begin with. Some details on this algorithm: https://www.geometrictools.com/Documentation/MinimumAreaRectangle.pdf
My main issue is in finding the vertices within the data that I currently have. I believe I need to at least find candidate vertices in order to reduce the amount of iterations I am performing, since the amount of points is relatively large and treating the interior points as if they are vertices is unnecessary if I can figure out a way to not include them.
Here is some example data that I am working with:
Here's the segmented scene using the naive algorithm, where it segments out the central objects relatively well due to the objects mostly being aligned with the image axes:
.
In red, you can see the current bounding boxes that I am drawing utilizing 2 vertices: top-left and bottom-right corners of the groups of points I have found.
The rotation part is where my current approach fails, as I am only defining the bounding box using two points, anything that is rotated and not axis-aligned will occupy much more area than necessary to encapsulate the points.
Here's an example with rotated objects in the scene:
Here's the current naive segmentation's performance on that scene, which is drawing larger than necessary boxes around the rotated objects:
Ideally the result would be bounding boxes aligned with the longest axis of the points that are being segmented, which is what I am having trouble implementing.
Here's an image roughly showing what I am really looking to accomplish:
You can also notice unnecessary segmentation done in the image around the borders as well as some small segments, which should be removed with some further heuristics that I have yet to develop. I would also be open to alternative segmentation algorithm suggestions that provide a more robust detection of the objects I am interested in.
I am not sure if this question will be completely clear, therefore I will try my best to clarify if it is not obvious what I am asking.
It's late, but that might still help. This is what you need to do:
expand pixels to make small segments connect larger bodies
find connected bodies
select a sample of pixels from each body
find the MBR ([oriented] minimum bounding rectangle) for selected set
For first step you can perform dilation. It's somehow like DBSCAN clustering. For step 3 you can simply select random pixels from a uniform distribution. Obviously the more pixels you keep, the more accurate the MBR will be. I tested this in MATLAB:
% import image as a matrix of 0s and 1s
oI = ~im2bw(rgb2gray(imread('vSb2r.png'))); % original image
% expand pixels
dI = imdilate(oI,strel('disk',4)); % dilated
% find connected bodies of pixels
CC = bwconncomp(dI);
L = labelmatrix(CC) .* uint8(oI); % labeled
% mark some random pixels
rI = rand(size(oI))<0.3;
sI = L.* uint8(rI) .* uint8(oI); % sampled
% find MBR for a set of connected pixels
for i=1:CC.NumObjects
[Y,X] = find(sI == i);
mbr(i) = getMBR( X, Y );
end
You can also remove some ineffective pixels using some more processing and morphological operations:
remove holes
find boundaries
find skeleton
In MATLAB:
I = imfill(I, 'holes');
I = bwmorph(I,'remove');
I = bwmorph(I,'skel');
The image below has many circles. Click and zoom in to see the circles.
https://drive.google.com/open?id=1ox3kiRX5hf2tHDptWfgcbMTAHKCDizSI
What I want is counting the circles using any free language, such as python.
Is there a function or idea to do it?
Edit: I came up with a better solution, partially inspired by this answer below. I thought of this method originally (as noted in the OP comments) but I decided against it. The original image was just not good enough quality for it. However I improved that method and it works brilliantly for the better quality image. The original approach is first, and then the new approach at the bottom.
First approach
So here's a general approach that seems to work well, but definitely just gives estimates. This assumes that circles are roughly the same size.
First, the image is mostly blue---so it seems reasonable to just do the analysis on the blue channel. Thresholding the blue channel, in this case, using Otsu thresholding (which determines an optimal threshold value without input) seems to work very well. This isn't too much of a surprise since the distribution of color values is pretty much binary. Check the mask that results from it!
Then, do a connected component analysis on the mask to get the area of each component (component = white blob in the mask). The statistics returned from connectedComponentsWithStats() give (among other things) the area, which is exactly what we need. Then we can simply count the circles by estimating how many circles fit in a given component based on its area. Also note that I'm taking the statistics for every label except the first one: this is the background label 0, and not any of the white blobs.
Now, how large in area is a single circle? It would be best to let the data tell us. So you could compute a histogram of all the areas, and since there are more single circles than anything else, there will be a high concentration around 250-270 pixels or so for the area. Or you could just take an average of all the areas between something like 50 and 350 which should also get you in a similar ballpark.
Really in this histogram you can see the demarcations between single circles, double circles, triple, and so on quite easily. Only the larger components will give pretty rough estimates. And in fact, the area doesn't seem to scale exactly linearly. Blobs of two circles are slightly larger than two single circles, and blobs of three are larger still than three single circles, and so on, so this makes it a little difficult to estimate nicely, but rounding should still keep us close. If you want you could include a small multiplication parameter that increases as the area increases to account for that, but that would be hard to quantify without going through the histogram analytically...so, I didn't worry about this.
A single circle area divided by the average single circle area should be close to 1. And the area of a 5-circle group divided by the average circle area should be close to 5. And this also means that small insignificant components, that are 1 or 10 or even 100 pixels in area, will not count towards the total since round(50/avg_circle_size) < 1/2, so those will round down to a count of 0. Thus I should just be able to take all the component areas, divide them by the average circle size, round, and get to a decent estimate by summing them all up.
import cv2
import numpy as np
img = cv2.imread('circles.png')
mask = cv2.threshold(img[:, :, 0], 255, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)[1]
stats = cv2.connectedComponentsWithStats(mask, 8)[2]
label_area = stats[1:, cv2.CC_STAT_AREA]
min_area, max_area = 50, 350 # min/max for a single circle
singular_mask = (min_area < label_area) & (label_area <= max_area)
circle_area = np.mean(label_area[singular_mask])
n_circles = int(np.sum(np.round(label_area / circle_area)))
print('Total circles:', n_circles)
This code is simple and effective for rough counts.
However, there are definitely some assumptions here about the groups of circles compared to a normal circle size, and there are issues where circles that are at the boundaries will not be counted correctly (these aren't well defined---a two circle blob that is half cut off will look more like one circle---no clear way to count or not count these with this method). Further I just used automatic thresholding via Otsu here; you could get (probably better) results with more careful color filtering. Additionally in the mask generated by Otsu, some circles that are masked have a few pixels removed from their center. Morphology could add these pixels back in, which would give you a (slightly larger) more accurate area for the single circle components. Either way, I just wanted to give the general idea towards how you could easily estimate this with minimal code.
New approach
Before, the goal was to count circles. This new approach instead counts the centers of the circles. The general idea is you threshold and then flood fill from a background pixel to fill in the background (flood fill works like the paint bucket tool in photo editing apps), that way you only see the centers, as shown in this answer below.
However, this relies on global thresholding, which isn't robust to local lighting changes. This means that since some centers are brighter/darker than others, you won't always get good results with a single threshold.
Here I've created an animation to show looping through different threshold values; watch as some centers appear and disappear at different times, meaning you get different counts depending on the threshold you choose (this is just a small patch of the image, it happens everywhere):
Notice that the first blob to appear in the top left actually disappears as the threshold increases. However, if we actually OR each frame together, then each detected pixel persists:
But now every single speck appears, so we should clean up the mask each frame so that we remove single pixels as they come (otherwise they may build up and be hard to remove later). Simple morphological opening with a small kernel will remove them:
Applied over the whole image, this method works incredibly well and finds almost every single cell. There are only three false positives (detected blob that's not a center) and two misses I can spot, and the code is very simple. The final thing to do after the mask has been created is simply count the components, minus one for the background. The only user input required here is a single point to flood fill from that is in the background (seed_pt in the code).
img = cv2.imread('circles.png', 0)
seed_pt = (25, 25)
fill_color = 0
mask = np.zeros_like(img)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3))
for th in range(60, 120):
prev_mask = mask.copy()
mask = cv2.threshold(img, th, 255, cv2.THRESH_BINARY)[1]
mask = cv2.floodFill(mask, None, seed_pt, fill_color)[1]
mask = cv2.bitwise_or(mask, prev_mask)
mask = cv2.morphologyEx(mask, cv2.MORPH_OPEN, kernel)
n_centers = cv2.connectedComponents(mask)[0] - 1
print('There are %d cells in the image.'%n_centers)
There are 874 cells in the image.
One possible solution would be to read the image using OpenCV, get its grayscale, then use Canny edge detection and perform countour finding in OpenCV. This will return a list of countours. It would look something like:
import cv2
image = cv2.imread('path-to-your-image')
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
# tweak the parameters of the GaussianBlur for best performance
blurred = cv2.GaussianBlur(gray, (7, 7), 0)
# again, try different values here
edged = cv2.Canny(blurred, 20, 140)
(_, contours, _) = cv2.findContours(edged.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
print(len(contours))
If you have all images like this - consider thresholding it, not necessarily by auto threshold-seeking algorithm like Otsu, but rather using simplest threshold by a given threshold value. Yes, before thresholding you have to convert your color input to gray-scale, or take one of color channels. Then based on few experiments with channels and threshold values - determine threshold value to have circles with holes in monochrome thresholding result. Based on your png image I found value of 81 (intensity of gray varies from 0 to 255) to be great to threshold gray-scale version of your input to have such binary image with holes in place, as described above.
Then simply count those holes.
Holes can be determined by seed-filling white area, connected to image border. As result you will have white hole connected components on black background - so simply count them.
More details you can find here http://www.leptonica.com/filling.html and use leptonica primitives to do thresholding, hole counting an so on.
I want to use FFT to accelerate 2D convolution. The filter is 15 x 15 and the image is 300 x 300. The filter's size is different with image so I can not doing dot product after FFT. So how to transform the filter before doing FFT so that its size can be matched with image?
I use the convention that N is kernel size.
Knowing the convolution is not defined (mathematically) on the edges (N//2 at each end of each dimension), you would loose N pixels in totals on each axis.
You need to make room for convolution : pad the image with enough "neutral values" so that the edge cases (junk values inserted there) disappear.
This would involve making your image a 307x307px image (with suitable padding values, see next paragraph), which after convolution gives back a 300x300 image.
Popular image processing libraries have this already embedded : when you ask for a convolution, you have extra arguments specifying the "mode".
Which values can we pad with ?
Stolen with no shame from Numpy's pad documentation
'constant' : Pads with a constant value.
'edge' : Pads with the edge values of array.
'linear_ramp' : Pads with the linear ramp between end_value and the arraydge value.
'maximum' :
Pads with the maximum value of all or part of the
vector along each axis.
'mean'
Pads with the mean value of all or part of the
vector along each axis.
'median'
Pads with the median value of all or part of the
vector along each axis.
'minimum'
Pads with the minimum value of all or part of the
vector along each axis.
'reflect'
Pads with the reflection of the vector mirrored on
the first and last values of the vector along each
axis.
'symmetric'
Pads with the reflection of the vector mirrored
along the edge of the array.
'wrap'
Pads with the wrap of the vector along the axis.
The first values are used to pad the end and the
end values are used to pad the beginning.
It's up to you, really, but the rule of thumb is "choose neutral values for the task at hand".
(For instance, padding with 0 when doing averaging makes little sense, because 0 is not neutral in an average of positive values)
it depends on the algorithm you use for the FFT, because most of them need to work with images of dyadic dimensions (power of 2).
Here is what you have to do:
Padding image: center your image into a bigger one with dyadic dimensions
Padding kernel: center you convolution kernel into an image with same dimensions as step 1.
FFT on the image from step 1
FFT on the kernel from step 2
Complex multiplication (Fourier space) of results from steps 3 and 4.
Inverse FFT on the resulting image on step 5
Unpadding on the resulting image from step 6
Put all 4 blocs into the right order.
If the algorithm you use does not need dyadic dimensions, then steps 1 is useless and 2 has to be a simple padding with the image dimensions.
Assuming that I have a grayscale (8-bit) image and assume that I have an integral image created from that same image.
Image resolution is 720x576. According to SURF algorithm, each octave is composed of 4 box filters, which are defined by the number of pixels on their side. The
first octave uses filters with 9x9, 15x15, 21x21 and 27x27 pixels. The
second octave uses filters with 15x15, 27x27, 39x39 and 51x51 pixels.The third octave uses filters with 27x27, 51x51, 75x75 and 99x99 pixels. If the image is sufficiently large and I guess 720x576 is big enough (right??!!), a fourth octave is added, 51x51, 99x99, 147x147 and 195x195. These
octaves partially overlap one another to improve the quality of the interpolated results.
// so, we have:
//
// 9x9 15x15 21x21 27x27
// 15x15 27x27 39x39 51x51
// 27x27 51x51 75x75 99x99
// 51x51 99x99 147x147 195x195
The questions are:What are the values in each of these filters? Should I hardcode these values, or should I calculate them? How exactly (numerically) to apply filters to the integral image?
Also, for calculating the Hessian determinant I found two approximations:
det(HessianApprox) = DxxDyy − (0.9Dxy)^2 anddet(HessianApprox) = DxxDyy − (0.81Dxy)^2Which one is correct?
(Dxx, Dyy, and Dxy are Gaussian second order derivatives).
I had to go back to the original paper to find the precise answers to your questions.
Some background first
SURF leverages a common Image Analysis approach for regions-of-interest detection that is called blob detection.
The typical approach for blob detection is a difference of Gaussians.
There are several reasons for this, the first one being to mimic what happens in the visual cortex of the human brains.
The drawback to difference of Gaussians (DoG) is the computation time that is too expensive to be applied to large image areas.
In order to bypass this issue, SURF takes a simple approach. A DoG is simply the computation of two Gaussian averages (or equivalently, apply a Gaussian blur) followed by taking their difference.
A quick-and-dirty approximation (not so dirty for small regions) is to approximate the Gaussian blur by a box blur.
A box blur is the average value of all the images values in a given rectangle. It can be computed efficiently via integral images.
Using integral images
Inside an integral image, each pixel value is the sum of all the pixels that were above it and on its left in the original image.
The top-left pixel value in the integral image is thus 0, and the bottom-rightmost pixel of the integral image has thus the sum of all the original pixels for value.
Then, you just need to remark that the box blur is equal to the sum of all the pixels inside a given rectangle (not originating in the top-lefmost pixel of the image) and apply the following simple geometric reasoning.
If you have a rectangle with corners ABCD (top left, top right, bottom left, bottom right), then the value of the box filter is given by:
boxFilter(ABCD) = A + D - B - C,
where A, B, C, D is a shortcut for IntegralImagePixelAt(A) (B, C, D respectively).
Integral images in SURF
SURF is not using box blurs of sizes 9x9, etc. directly.
What it uses instead is several orders of Gaussian derivatives, or Haar-like features.
Let's take an example. Suppose you are to compute the 9x9 filters output. This corresponds to a given sigma, hence a fixed scale/octave.
The sigma being fixed, you center your 9x9 window on the pixel of interest. Then, you compute the output of the 2nd order Gaussian derivative in each direction (horizontal, vertical, diagonal). The Fig. 1 in the paper gives you an illustration of the vertical and diagonal filters.
The Hessian determinant
There is a factor to take into account the scale differences. Let's believe the paper that the determinant is equal to:
Det = DxxDyy - (0.9 * Dxy)^2.
Finally, the determinant is given by: Det = DxxDyy - 0.81*Dxy^2.
Look at page 17 of this document
http://www.sci.utah.edu/~fletcher/CS7960/slides/Scott.pdf
If you made a code for normal Gaussian 2D convolution, just use the box filter as a Gaussian kernel and the input image will be the same original image not integral image. The results from this method will be same with the one you asked.
I study convolution in image processing as it is a part of the curriculum, I understand the theory and the formula but I am confused about its implementation.
The formula is:
What I understand
The convolution kernel is flipped both horizontally and vertically then the values in the kernel are multiplied by the corresponding pixel values, the results are summed, divided by "row x column" to get the average, and then finally this result is the value of the pixel at the center of the kernel location.
Confusion in implementation
When I run the example convolution program from my course material and insert as input a 3x3 convolution kernel where:
1st row: (0, 1, 0)
2nd row: (0, 0, 0)
3rd row: (0, 0, 0)
The processed image is shifted down by one pixel, where I expected it to shift upwards by one pixel. This result indicates that no horizontal or vertical flipping is done before calculating (as if it is doing correlation).
I thought there might be a fault in the program so I looked around and found that Adobe Flex 3 and Gimp are doing this as well.
I don't understand, is there something that I missed to notice?
Appreciate any help or feedback.
I guess the programs you tried implement correlation instead of convolution.
I've tried your filter in Mathematica using the ImageFilter function, the result is shifted upwards as expected:
result:
I've also tried it in Octave (an open source Matlab clone):
imfilter([1,1,1,1,1;
2,2,2,2,2;
3,3,3,3,3;
4,4,4,4,4;
5,5,5,5,5],
[0,1,0;
0,0,0;
0,0,0],"conv")
("conv" means convolution - imfilter's default is correlation). Result:
2 2 2 2 2
3 3 3 3 3
4 4 4 4 4
5 5 5 5 5
0 0 0 0 0
Note that the last row is different. That's because different implementations use different padding (by default). Mathematica uses constant padding for ImageConvolve, no padding for ListConvolve. Octave's imfilter uses zero padding.
Also note that (as belisarius mentioned) the result of a convolution can be smaller, same size or larger than the source image. (I've read the terms "valid", "same size" and "full" convolution in the Matlab and IPPI documentation, but I'm not sure if that's standard terminology). The idea is that the summation can either be performed
only over the source image pixels where the kernel is completely inside the image. In that case, the result is smaller than the source image.
over every source pixel. In that case, the result has the same size as the source image. This requires padding at the borders
over every pixel where any part of the kernel is inside the source image. In that case, the result image is larger than the source image. This also requires padding at the borders.
Please note that:
Results in:
So, the "shifting" is not real, as the dimensions are affected.