I'm working on MATLAB on some regions inside an image. I'm at a point in which I would like to be able to separate regions which exhibit some kind of regularity (e.g., being circle-ish or square-ish) from regions which does not resemble any known figure and which for my application are mere noise. I'll illustrate this using a descriptive MS Paint image:
Is there any tool that, most of the times (or even less, I know this can't be 100/100) will recognize the red thing as being different?
I'll deal with many shapes in a single image, so I don't mind if I carry on some red monsters along the way, as long as the majority of them is kicked out. Of course I know the indices of these regions, so I can manipulate them in MATLAB.
Many algorithms come to mind, e.g., getting the boundary and checking for its regularity/the number of times it changes curvature/..., checking for variations in vertical length through different columns (nearly 0 for the linear feature, really high for the red stuff), ...
However I was hoping in some help from a tool out there. It doesn't matter if this tool won't cover all cases (for example, will kick out circles), I've been very broad to get the maximum number of inputs from you guys - any tool will be inspiring and helpful (and, however, we can't expect a perfect answer for the deeper question - recognizing regular shapes - which seems more like a AI field of research). I also think that, while being broad, this is totally non-subjective so should fit in SO. Thank you.
Side note 1: I'll deal mostly with elongated, extended features like the top-right one, so circles are not that relevant.
Side note 2: To be 100% clear, I would need something (be it an already existant tool, or some ideas pointed out by you) that acts on the indices of the shapes, in terms of rows-columns into the original image, or on the boundary of the shape itself.
Side note 3: Apart from tools/suggestions/ideas, you are welcomed to write down some lines of code ;) I'm getting the regions as connected components from bwconncomp.
I had to solve a similar problem recently that involved counting the number of indentations on blobs within in an image (basically, the connected components returned by bwconncomp). The method I used was to look at curvature changes along the boundary calculated via the FFT. In your case, the red blobs would have a large number of curvature variations, whereas the black regions would not. It's a pretty easy calculation and relatively fast. The code is on github here:
https://github.com/mjsottile/blobdents
The file of interest is src/countindents.m. A short description of the approach is here:
http://arxiv.org/abs/1501.07692
I went for the easier road as suggested by #Mikhail in comments.
I found out regionprops has a really helpful tool called Solidity. Quoting docs,
Returns a scalar specifying the proportion of the pixels in the convex hull that are also in the region. Computed as Area/ConvexArea.
Convex hull is defined as the smallest convex polygon that can contain the region. So Solidity goes up to 1 if the shape is kind of regular and has no convexity changes; down to 0 for my red shape, which leaves space between itself and the convex polygon.
Of course it never reaches 0, lowest value should belong to a kind of +-shaped sign.
Related
Which method is commonly used to evaluate the remaining 'boundary' pixels after an initial segmentation (based on thresholds)?
I thought about classification based on a standard deviation from the threshold values but I don't know if that is common practice in image analysis. This would be a region growing method but based on the answer on this question ( http://www.mathworks.com/matlabcentral/answers/53351-how-can-i-segment-a-color-image-with-region-growing ) it is not sensible to use the region growing algorithm. Someone suggested imdilate. This method seems arbitrary, useful when enhancing images for aesthetic purpose or to enhance the visibility. For my problem the assigning of the pixels has to be correct because I have to do measurements on these extracted objects/features and a few pixels make a huge difference.
What I was looking for :
To collect my boundary pixels of the BW image from the first segmentation (which I found : http://nl.mathworks.com/help/images/ref/bwboundaries.html)
A decision rule (nearest neighbor ?) to classify those boundary pixels. It would be helpful if there were multiple methods to do this, because it makes a relative accuracy check of the classification possible.
I would really appreciate the input/advice from someone with more experience in this area to point me to the right direction (functions, tutorials etc…)
Thank you !
What will work for you depends very much on the images you have. This is no one-size-fits-all algorithm.
First, you need to answer the question: Given a pixel close to a segmented feature, what would make you believe that this pixel belongs to the feature? Also: what is "close"?
The answer to the second question determines your search area. Here, imdilate is useful to identify candidate pixels (i.e. you dilate your feature, subtract the feature, and you are left with a ring of candidate pixels around each feature). If you test on all pixels, the risk is not so much that it could take forever, but that for some images, your region growing mechanism expands to the entire image.
The answer to the first question determines what algorithm you'll use. Do you look for a gradient, i.e. "if pixel p is closer in intensity to the adjacent feature than to most of its neighbors, then I take it"? Do you look for texture? Do you look for a local threshold (hysteresis thresholding)? The answer, again, depends very much on the images you are segmenting. Make sure you test on a large set of images, because what may look good on one image may totally fail on a different one.
I have two sets of 2D Points (shown in images below).
And I would like to find some high confidence correspondence between these dots.
These dots are extracted feature points from 2 camera images from different angles. Two images are relatively well rectified, though not perfect. However, there will be distortion/warp caused by depth in the scene, the number of points might not be the same, there might be outliers, etc.
One approach could be using a sliding window that contains multiple dots and try block matching. But that might be kind of slow. I feel like there should be a relatively straight forward solution to this problem.
For example, this paper might be addressing a similar problem.
You can use each dot/point in one of the images, and search for its "neighbors" in the other image.
Just a few days ago someone asked a similar question here, and got a very sophisticated (accepted) answer:
How to calculate the nearest neighbors using weka from the command line?
But maybe your problem is so common in image processing that there are even better solutions, but you might try this one (implemented in java).
Given a set of non-rotated AABB bounds, I'm hoping to create a simpler set of bounds from the original set, that allows for a specified amount of inaccuracy.
Some examples:
I'm working with this in Unity with Bounds, but it's just basic AABB comparison stuff, nothing Unity-specific. I figure someone must have worked out a system for this at some point in the past, but I had no luck searching around. Encapsulating bounds are easy but this is harder, since you can't just iterate through each bounds one by one. Sometimes a simpler solution can only be seen by looking at the whole thing.
Fast performance isn't critical but would be nice. Inaccuracy is OK in both directions (i.e. the bounds may cover a little less than the actual size or a little more). If it helps, I can expect all bounds in the original set to be connected somewhere - no free-floating pieces in a separate group.
I don't expect anyone to write up a whole system to solve this, I'm more hoping that it's already been solved or that maybe there's an obvious process to achieve it that I haven't thought of yet.
This sounds something that could be handled with Surface Area Heuristics (SAH). SAH is commonly used in ray tracing to build better tree like structures were the triangles are stored. There are multiple sources discussing it more. One good is Wald's thesis chapter 7.3.
The basic idea in the SAH built is to start with the whole space and divide it recursively. Division position is decided by sweeping through all reasonable positions and calculating surface area of both child nodes. The reasonable positions are the positions were any triangle has its upper or lower bound. After sweeping through all the candidates, the division with the smallest total surface area in the children is used.
If SAH is not a good idea for your application, you could use similar sweeping through all candidates, but calculate for example the extra space inside the AABBs.
I am going through feature detection algorithms and a lot of things seems to be unclear. The original paper is quite complicated to understand for beginners in image processing. Shall be glad if these are answered
What are the features which are being detected by SURF and SIFT?
Is it necessary that these have to be computed on gray scale images?
What does the term "descriptor" mean in simple words.
Generally,how many features are selected/extracted?Is there a criteria for that?
What does the size of Hessian matrix determine?
What is the size of the features being detected?It is said that the size of a feature is the size of the blob.So, if size of image is M*N so will there be M*N n umber of features?
These questions may seem too trivial, but please help..
I will try to give an intuitive answer to some of your questions, I don't know answers to all.
(You didn't specify which paper you are reading)
What are the features and how many features are being detected by SURF
and SIFT?
Normally features are any part in an image around which you selected a small block. You move that block by a small distance in all directions. If you find considerable variations between the one you selected and its surroundings, it is considered as a feature. Suppose you moved your camera a little bit to take the image, still you will detect this feature. That is their importance. Normally best example of such a feature is corners in the image. Even edges are not so good features. When you move your block along the edge lines, you don't find any variation, right?
Check this image to understand what I said , only at the corner you get considerable variation while moving the patches, in other two cases you won't get much.
Image link : http://www.mathworks.in/help/images/analyzing-images.html
A very good explanation is given here : http://aishack.in/tutorials/features-what-are-they/
This the basic idea and the algorithms you mentioned make this more robust to several variations and solve many issues. (You can refer their papers for more details)
Is it necessary that these have to be computed on gray scale images?
I think so. Anyway OpenCV works on grayscale images
What does the term "descriptor" mean in simple words?
Suppose you found features in one image, say image of a building. Now you took another image of same building but from a slightly different direction. You found features in the second image also. But how can you match these features. Say feature 1 in image 1 match to which feature in image 2 ? (As a human, you can do easily, right ? This corner of building in first image corresponds to this corner in second image, so and so. Very easy).
Feature is just giving you pixel location. You need more information about that point to match it with others. So you have to describe the feature. And this description is called "descriptors". To describe this features, algorithms are there and you can see it SIFT paper.
Check this link also : http://aishack.in/tutorials/sift-scale-invariant-feature-transform-introduction/
Generally,how many features are selected/extracted?Is there a criteria
for that?
During processing you can see applying different thresholds, removing weak keypoints etc. It is all part of plan. You need to understand algorithm to understand these things. Yes, you can specify these threshold and other parameters (in OpenCV) or you can leave it as default. If you check for SIFT in OpenCV docs, you can see function parameters to specify number of features, number of octave layers, edge threshold etc.
What does the size of Hessian matrix determine?
That I don't know exactly, just it is a threshold for keypoint detector. Check OpenCV docs : http://docs.opencv.org/modules/nonfree/doc/feature_detection.html#double%20hessianThreshold
I've done work on software used for controlling imaging hardware, such as microscopes, that are sometimes hard to get time on. This means it is difficult to test out new/different algorithms which would require access to the instrument. I'd like to create a synthetic instrument that could be used for some of these testing purposes, and I was thinking of using some kind of fractal image generation to create the synthetic images. The key would be to be able to generate features at many different 'magnifications' and locations in some sort of deterministic manner. This is because some of the algorithms being tested may need to pan/zoom and relocate previously 'imaged' areas. Onto these base images I can then apply whatever instrument 'defects' are appropriate (focus, noise, saturation, etc.).
I'm at a bit of a loss on how to select/implement a good fractal algorithm for the base image. Any help would be appreciated. Preferably it would have the following qualities:
Be fast at rendering new image areas.
Fairly wide 'feature' coverage at as many locations and scales as possible.
Be deterministic (but initialized from random starting parameters).
Ability to tune to make images look more like 'real' images.
Item 2 is important, for example a mandelbrot set, with its large smooth/empty regions, might not be good since the software controlling the synthetic scope might fall into one of these areas.
So far I've thought of using something like a mandelbrot, but randomly shifting/rotating/scaling and merging two or more fractal sets to get more complete 'feature' coverage.
I've also seen images of the fractal flame algorithms and they seem to generate images that might be useful (and nice to look at).
Finally, I've thought of using some sort of paused particle simulation run to generate images that are more cell-like (my current imaging target), but I'm not sure if this approach can be made to work with the other requirements.
Edit:
#Jeffrey - So it sounds like some kind of terrain generation might be the way to go, as long as I have complete control over the PSRNG. Perhaps I can use some stored initial seed + x position + y position to generate my random numbers? But then I am unsure of how to consistently generate the terrains across scales, except, as you mentioned, to create the base terrain at the coursest scale, and at certain pre-determined 'magnifications' add new deterministic pseudo-random variations to this base. I'd also have to be careful about when to generate the next level of terrain, since if I'm too aggressive I'd have to generate and integrate the results appropriately for display at the coarser level... This is why I initially was leaning toward a more 'traditional' fractal, since this integration from finer scales would be handled more implicitly (I think).
The idea behind a fractal terrain creation algorithm is to build the image at each scale separately. For a landscape it's easy: just make a small array of height values, and set them randomly. Then scale it up to a larger array, averaging the values so that the contour is smooth, and then add small random amounts to those values. Then scale it up, etc. The original small bumps have become mountains, and they are filled with complex terrain.
There are two particular difficulties with the problem posed here, though. First, you don't want to store any of these values, since it would be potentially huge. Secondly, the features at each scale are of a different kind than the features at other scales.
These problems are not insurmountable.
Basically, you would divide the image up into a grid, and using deterministic psedorandom numbers establish the key features of each square in the grid. For example, each square could have a certain density of cell types.
At the next level of magnification, subdivide each square into another grid, apply a gradiant of values across the grid that is based on the values of the containing square and its surrounding squares. Then apply pseudorandom variations to that seeded with the containing square's grid coordinates. For the random seed, always use the coordinates of the immediately containing square of the subdivision under consideration regardless of where the image is cropped, in order to ensure that it is recreated correctly accross multiple runs.
At some level of magnification the random values go from being densities of paticles types to particle locations. Then for each particle, there are partical features. Then features on those features.
Although arbitrary left/right and up/down scrolling will be desired, the image at all levels of magnification above the current scene will have to be calculated each time the frame is shifted to ensure that all necessary features are included. This way the image can be scrolled from one cell to another without loss of consistancy. Partical simulations can be used to ensure that cells or cell features don't overlap. This could be done in a repeatable, deterministic manner.
And don't forget to apply a smoothing gradient based on averages of surrounding squares at higher levels before adding in the random variations. Otherwise, the abrupt changes will make the squares themselves appear in the images!
This answer is somewhat rambling and probably confusing, but that is best I can explain it right now. I hope it helps!