I would like to apply OCR to some pictures of 7 segment displays on a wall. My strategy is the following:
Covert Img to Grayscale
Blur img to reduce false edges
Threshold the img to a binary img
Apply Canny Edge detection
Set Region of Interest (ROI) base on a pattern given by the silhouette of the number
Scale ROI and Template match the region
How to set a ROI so that my program doesn't have to look for the template through the whole image? I would like to set my ROI base on the number of edges found or something more useful if someone can help me.
I was looking into Cascade Classification and Haar but I don't know how to apply it to my problem.
Here is an image after being pre-processed and edge detected:
original Image
If this is representative of the number of edges you'll have to deal with you could try a nice naive strategy like sliding a ROI-finder window across the binary image which just sums the pixel values, and doesn't fire unless that value is above a threshold. That should optimise out all the blank surfaces.
Edit:
Ok some less naive approaches. If you have some a-priori knowledge, like you know the photo is well aligned (and not badly rotated or skewed), you could do some passes with a low-high-low-high grate tuned to capture the edges either side of a segment, using different scales in both x and y dimensions. A good hit in both directions will give clues not only about ROI but what scale of template to begin with (too large and too small grates won't hit both edges at once).
You could do blob detection, and then apply your templates to blobs in turn (falling back on merging blobs if the template matching score is below a threshold, in case your number segment is accidentally partitioned). The size of the blob might again give you some hint as to the scale of template to apply.
First of all, given that the original image has a LED display and so the illuminated region is has a higher intensity than the trest, I'd perform say a Yuv colour transformation on the original image and then work with the intensity plane (Y).
Next, if you know that the image is well aligned (i.e. not rotated) I would suggest applying separate horizontal and vertical edge detectors rather than a generic edge detector (you are not interested in diagonal lines). E.g.
sobelx = cv2.Sobel( img, cv2.CV_64F, 1, 0, ksize=5 )
sobely = cv2.Sobel( img, cv2.CV_64F, 0, 1, ksize=5 )
Otherwise you might use contour detection to find the bounds of the digits (though you may need to perform a dilate to close the gaps between LED segments.
Next I would construct horizontal and vertical histograms of the output from these edge or contour detections. These would help you to identify 'busy' regions of the image which contain many edges.
Finally, I'd threshold the Y plane and explore each of the ROIs with my template.
Related
I am new to opencv, coding in c++. I have a task given to me to decode a 2D circle barcode using an encoded array. I am up to the point where I am able to centralize the figure and get the line using Hough transforms.
Need help with how to read the colour in the images, note that each of the two adjacent blocks correspond to a letter.
Any pointers will be highly appreciated. Thanks.
First, you need to load the image. I suspect this isn't a problem because you are already using Hough transforms on it, but:
Mat img = imread(filename)
Once the image is loaded, you can grab any of the pixels using:
Scalar intensity = img.at<uchar>(y, x);
However, what you need to do is threshold the image. As I mentioned in the comments, the image colors are either 0 or 255 for each RGB channel. This is on purpose for encoding the data in case there are image artifacts. If the channel is above a certain color value, then you will consider that it's 'on' and if below, it's 'off'.
Threshold the image using adaptiveThreshold. I would threshold down to binary 1 or 0. This will produce RGB triplets that are one of eight (2^3) possible combinations, from (0,0,0) to (1,1,1).
Then you need to walk the pixels. This is where it gets interesting.
You say each adjacent 2 pixels form a single letter. That's 2^6 or 64 different letters. The next question is: are the letters arranged in scan lines, left-to-right, top to bottom? If yes, then it will be important to orientate the image using the crosshair in the center.
If the image is encoded radially (using polar coordinates) then things get a little trickier. You need to use cvLinearPolar to remap the image.
Otherwise you need to walk the whole image, stepping the size of the RGB blocks and discard any pixels whose distance from the center is greater than the radius of the circle. After reading all of the pixels into an array, group them by pairs.
At some point, I would say that using OpenCV to do this is heading towards machine learning. There has to be some point where you can cut in and use Neural Networks to decode the image for you. Once you have the circle (cutoff radius) and the image centered, you can convert to polar coordinates and discard everything outside the circle by cutting off everything greater than the radius of the circle. Remember, polar coordinates are (r,theta), so you should be able to cutoff the right part of the polar image.
Then you could train a Neural Network to take the polar image as input and spit out the paragraph.
You would have to provide lots of training data, and the trained model would still be reliant on your ability to pre-process the image. This will include any affine transforms in case the image is tilted or rotated. At that point you would say to yourself that you've done all the heavy lifting and the last little bit really isn't that hard.
However, once you get a process working for a clean image, you can start adding to steps to introduce ML to work on dirty images. HoughCircles can be used to detect the part of an image to run detection on. Next, you need to decide if the image inside the circle is a barcode or not.
A good barcode system will have parity bits or some other form of error correction, but you can use machine learning to cleanup output.
My 2 cents anyways.
I have this raw image in grayscale:
I would like to detect the edge of the object. However, it is affected by illumination near the edge. This is what I obtained after Gaussian blur and Canny edge detection:
This is my code:
cv::cvtColor(imgOriginal, imgGrayscale, CV_BGR2GRAY); // convert to grayscale
cv::GaussianBlur(crop, // input image
imgBlurred, // output image
cv::Size(5, 5), // smoothing window width and height in pixels
5); // sigma value, determines how much the image will be blurred
cv::Canny(imgBlurred, // input image
imgCanny, // output image
0, // low threshold
100); // high threshold
The light source is beneath the object. The illumination at the edge of object is from the light source or reflection of light. They are always at the same place.
The illumination is detected as edge as well. I have tried several other ways such as connected component labelling and binarize image with sample code (a beginner here) but to avail. Is there any way to detect clean edge illumination?
The background light patches may be removeable using some erosion with a fairly large kernel, since the object is much bigger than the light patches
Another common technique you can try is using distance transform + watershed. The distance transform will likely return points that you are certain are inside the object (since the object has little dark areas). Watershed will try to find regions that are connected (by comparing gradients) to the confirmed points. You may need to combine multiple regions after the watershed if the distance transform gives multiple points inside the object.
It is impossible to completely get rid of this problem. What an edge detector can detect are intensity variations that result due to edges in the objects. Given the lighting you have there, the variations caused by lighting are quite prominent.
I would suggest two approaches to solve this problem:
Adjust the lighting, if you are able to
Getting the lighting right solves 50% of any computer vision problem.
Use any knowledge that you have about the image, background or lighting to remove unnecessary edges. If the camera is stationary, background subtraction can remove edges resulting from background. If you know the shape, color, etc. of the object, you can remove edges that are not a good fit with the object. If it is too hard to determine the exact properties of the object, you can also train a machine learning system with many photos, to segment the image.
I want to detect multiple (similar) rectangular objects in an image that have a lot of substructure within them. So, my current plan is to use gaussian blur, morphology and edge detection (Canny). After using edge detection I get this (with very low threshold parameters):
What I want in the end is the outline of the greater rectangles. See:
Currently I try to get this by using HoughLines and findContours afterwards. For this to work, I need to fiddle a lot with the threshold parameters for Canny and HoughLines.
When I get it right for one image the parameters most likely will not work for the next one (e.g. the edges in the previous image were less dominant leading to too many lines detected by the hough transformation). Another problem is that sometimes inner structures are equally or less dominant than one side of the outer edges.
I tried to use a stronger blur or morphology but at some point this blurred away the small gap between the rectangles.
Can I extract the bigger rectangles somehow else given the edge image (preferably with opencv)?
Getting the 4 corner points would be enough.
I am trying to detect a large number of small circles that are in relatively close proximity to one another (only about 20 pixels apart) using OpenCV. I have managed to create this mask using cv::inRange() and cv::Canny().
Original Image
Mask
However, when I use cv::HoughCircles() only some of the circles are being detected accurately. Currently, I am using cv::HoughCircles() with the following parameters:
cv::HoughCircles(mat, circles, CV_HOUGH_GRADIENT, 2, mat.rows / 256, 100, 8, 2, 8);
Is this method not effective enough to detect circles that are this small and close together, or do I simply need to modify the parameters of cv::HoughCircles()?
Also, it would be useful to get rid of the "noise" surrounding the array of circles in the middle of the mask because some "false circles" are being detected around the edges of the mask. Is there a simple way to do this?
Get rid of the noise :
If you can make sure to always have the same environment parameters (e.g. distance from the circle, luminosity...), then you could mask your image just after the Canny edge detection, with cvAnd; here is what the mask would look like :
Hough circles detection :
Now, about HoughCircle. First, this function performs its own Canny edge detection. You are doing one too just before the call to HoughCircle. It may have an impact on the shapes of your circles, because of the way Canny works (i.e. intensity gradient on binary image...).
Speaking about the shape of your circles, just below is a close-up of what your "circles" look like; I would have been very impressed if HoughCircle actually did detect all or even just some of those. It can't give anything good in Hough space. Just to make sure, set the last two parameters to 0 (min/max radius), and try to lower the minimum distance between centers. But honestly, I think you need to find another approach to your problem.
[EDIT]
A possible approach would be to perform connected component labeling (e.g. blob detection). As far as I know it is not possible to do this simply with OpenCV alone, you will need something like cvblob, which is a very good OpenCV-based blob library. In particular, you might be interested in cvCentroid(CvBlob *blob).
Cheers
Hum, do you really need to detect them as circles? (as opposed to model them as circles).
If this is some kind of calibration pattern, and you are only interested in estimating the image positions of the centers, It may be a lot more efficient to detect them as point-like features first, then process each detected one individually - e.g. fitting a circle to a blob of white pixels in the neighborhood of each detected feature.
i have several contours that consist of several black regions in my image. Directly adjacent to these black regions are some brighter regions that do not belong to my contours. I want to add these brighter regions to my black region and therefor extend my contour in OpenCv.
Is there a convenient way to extend a contour? I thought about looking at intensity change from my gradient-image created with cv::Sobel and extend until the gradient changes again, meaning the intensity of pixel is going back to the neither black nor bright regions of the image.
Thanks!
Here are example images. The first picture shows the raw Image, the second the extracted Contour using Canny & findContours, the last one the Sobel-Gradient intensity Image of the same area.
I want to include the bright boundaries in the first image to the Contour.
Update: Now i've used some morphological operations on the Sobelgradients and added a contour around them (see Image below). Next step could be to find the adjacent pair of purple & red contours, but it seems very much like a waste of procession time to actually have to search for directly adjacent contours. Any better ideas?
Update 2: My solution for now is to search for morphed gradient (red) contours in a bounding box around my (purple) contours and pick the one with correct orientation & size. This works for gradient contours where the morphological operation closes the "rise" and "fall" gradient areas like in Figure 3. But it is still a bad solution for cases in which the lighted area is wider then in the image above. Any idea is still very much appreciated, thanks!
What you're trying to do is find two different features and merge them. It's not terribly difficult but you have to use multiple copies of the image to make it happen.
Make one copy, and threshold it for the dark portion
Make another copy and threshold it for the light portion
Merge both thresholded images into a new image
Apply a morphological operation like opening or closing (depending on how you threshold) This will connect nearby components
Find contours in the resultant image
Use those contours on your original image. This will work since all the images are the same size and all based off of the original.