I am doing a project on face recognition. I have a dataset containing image of 21 actors(each 150). Now I want to increase the no. of image of each actor to 300+ for the training purpose. How can I do it using MATLAB. Some solutions can be we can vary the contrast/brightness level of each image. But what are some other factors through which I can increase the no. of images.
One option is for you to flip the images: If a person is looking to the right, after the flip he will be looking to the left.
Further more, depending on your possible toolkit and set of skills, you could do some more advanced technique. If you can find some interesting characteristic from the pictures, like: eyes, nose, mouth, background. With those, you could make some intelligent transformations - swap peoples eyes, change the background, switch noses.
There are some particular objects of the faces which you could also distort - like the eyes and nose - stretch them. Maybe for bold guys you could built some synthetic hair, and so on...
You could do the contrast/brightness level change, but usually it doesn't do so well, as your features probably doesn't have (almost) anything to do with it, so it will just be a duplication of your data.
Anyway, as it's not a very large data set, if you don't have the set of skills to pull the more advances options I proposed, or the time to deal with it, you can make some of those stuff manually. It won't take you as much as you think. And usually, with that amount of data, this will give a good boost to your results.
What you are looking for is called "data augmentation". Common transformations are mirroring (flipping left / right side of the image) and rotation of the image. You might also be able to zoom (crop) a part of the image.
Maybe scaled versions with the rotated ones may help. If your features are not robust to the changes such as lightning contras etc you can modfy the images accordingly.
Related
I am capturing static images of particulate biological materials on the millimeter scale, and then processing them in MATLAB. My routine is working well so far, but I am using a rudimentary calibration procedure where I include some coins in the image, automatically find them based on their size and circularity, count their pixels, and then remove them. This allows me to generate a calibration line with input "area-mm^2" and output "Area- pixels," which I then use to convert the pixel area of the particles into physical units of millimeters squared.
My question is: is there a better calibrant object that I can use, such as a stage graticule or "phantom" as some people seem to call them? Do you know where I could purchase such a thing? I can't even seem to find a possible vendor. Is there another rigorous way to approach this problem without using calibrant objects in the field of view?
Thanks in advance.
Clay
Image calibration is always done using features of knowns size or distance.
You could calculate the scale based on nominal specifications but your imaging equipment will always have some production tolerances, your object distance is only known to a certain accuracy...
So it's always safer and simpler to actually calibrate your scale.
As a calibrant you can use anything that meets your requirements. If you know the size well enough and if you are able to extract it's dimensions in pixels properly you can use it...
I don't know your requirements and your budget, but if you want something very precise and fancy you can use glass masks.
There are temperature stable glass slides that are coated with chrome for example. There are many companies that produce such masks customized (IMT AG, BVM maskshop, ...) Also most optics lab equipment suppliers have such things on stock. Edmund Optics, Newport, ...
I was wondering if someone can provide me a guideline to detect if a person in a picture is bald or not, or even better, how much hair s\he has.
So far I tried to detect the face and the eyes position. From that information, I roughly estimate the forehead and bald area by cutting the area above the eyes as high as some portion of the face.
Then I extract HOG features and train the system with bald and not-bald images using SVM.
Now when I'm looking at the test results, I see some pictures classified as bald but some of them actually have blonde hair or long forehead that hair is not visible after the cutting process. I'm using MATLAB for these operations.
So I know the method seems to be a bit naive, but can you suggest a way of finding out the bald area or extracting the hair, if exists. What method would be the most appropriate for that kind of problem?
very general, so answer is general unless further info provided
Use Computer Vision (e.g MATLAB Computer Vision toolkit) to detect face/head
head has analogies (for human faces), using these one can get the area of the head where hair or baldness is (it seems you already have these)
Calculate the (probabilistic color space model) range where the skin of the person lies (most peorple have similar skin collor space range)
Calculate percentage of skin versus other color (meaning hair) in that area
You have it!
To estimate a skin color model check following papers:
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.56.8637&rep=rep1&type=pdf
http://infoscience.epfl.ch/record/135966
http://www.eurasip.org/Proceedings/Eusipco/Eusipco2010/Contents/papers/1569293757.pdf
Link
If an area does not fit well with skin model it can be taken as non-skin (meaning hair, assuming no hats etc are present in samples)
Head region is very small, hence, using HOG for classification doesn't make much sense.
You can use prior information - like detect faces; baldness/hair is certain to be found on the area above the face. Also, use some denser feature descriptors.
You are probably ending up with very sparse representation or equivalently less information because of which your classifier is not able to classify correctly.
I'm making an image processing project which has a 6-step algorithm and I'm stuck in one of these.
First off all, the platform I using is MATLAB, so if you can supply some samples it would be great. But if you don't want to write code samples, please just give me some hints, techniques or etc.
Let me explain my problem. I've segmented a .jpg image and cut out some areas of it. Then I save the result as .png using a mask. The result is like that (The black part is actually transparent, I made it black to see better the problem);
As you may see in the picture, there is some irrelevant areas. I need to get rid of these irrelevant areas. Because I want the foreground as much as smooth. At the first sight, I applied Gaussian blur to the mask and save the image as .png, again. But the result isn't satisfying as you can imagine. I suppose this situation is needed more solid solution than I have tried.
Edit1: I used spectral matting. But it doesn't help. The best result that I can receive is like that;
As you can see there is some problems on the face and lots of problems on the bottom side of the picture. I guess I need kind of edge fixer or edge smoother for the first image above and it should be faster than matting.
Any MATLAB code samples, technique and approach would be great. If you need further explanation, feel free to ask.
You do not want just to "Gauss-blur" the result, you want soft segmentation a.k.a matting. As a first stop for image matting I would recommend Levine Rav-Acha and Lischinski Spectral Matting. You'll find some Matlab code there (I used it in the past - very impressive results).
I have a question regarding computer vision; seems to be a general question but anyways, just wondering if you might have a clue. I was wondering if there is an efficient way to distinguish texture images (or photos with repetitive patterns) between whatnot, say realistic photos? The patterns could have exact repetitions, or just have major similarity. Actually I'm trying to see given an image if, it is possible to detect it is a texture or a pattern-based image, and that in real-time maybe?
For instance these three are considered textures in our context:
http://www.bigchrisart.com/sites/default/files/video/TR_Texture_RockWall.jpg
http://www.colourbox.com/preview/4440275-144135-seamless-geometric-op-art-texture.jpg
Thank you!
I cannot open your first image. I implemented the Fourier transform on your second one, and you can see frequency responses at specific points:
You can further process the image by extract the local maximum of the magnitude, and they share the same distance to the center (zero frequency). This may be considered as repetitive patterns.
Regarding the case that patterns share major similarity instead of repetitive feature, it is hard to tell whether the frequency magnitude still has such evident response. It depends on how the pattern looks like.
Another possible approach is the auto-correlation on your image.
I need to create an algorithm to layout some hierarchical data but have never done this kind of thing before and need some broad tips.
Basically I need to recreate this diagram (with dynamic data):
diagram http://dl.dropbox.com/u/15126868/diagram.png
bigger
I don't have a problem with most of it but need help with two things:
How do I approach writing a layout algorithm?
Should I use UIView subclasses for all discs or use quartz (I do need interaction)
Any suggestions most welcome. I don't need too much detail.
A bit more detail:
I'm currently thinking I should use UIView subclasses and layoutSubviews. Trouble is I need to know the size (at least roughly) of all nodes before I can start to position them. Then, as the positioning involves rotation, I may need to adjust child positioning again - and I can't add labels until after any rotation.
Other considerations seem to be: that the presentation area is rectangular, not square; that I can't spill off the page; and that I will need to animate changes to the sizes of the discs.
Any pointers would be great, thanks.
This sort of thing is very difficult.
Interestingly the perhaps main actual initial constraint here is the size of the typography.
In the example given: Observe they could have chosen a different SCPT** somewhat larger (perhaps, 10%-15% larger) or somewhat smaller and it would have still worked. They made an aesthetic decision on the SCPT.
White space is critical to design. Their particular graphic designer happened to like the particular feel of white space which you see. But it would have by no means been "wrong" with a smaller SCTP. Further, observe they could have used an even larger SCPT ... IF ... they used a smaller point size on the typography.
Note that in any event you simply won't be able to display that much type that small on an iPad (or Fone4).
So straight away you have to make decisions about how the type will appear, popup, audio or whatever. Even the white type ("on the discs" type) will give you trouble.
You will have to do lots of tests with photoshop first on to your iPad before even proceeding with an algorithm. So purely for what it's worth...
Here's how I personally would do this sort of thing. In general plan: I would try to do a squishy algorithm that retries itself until it finds a result it is happy with.
IMHO, based on previously doing this type of thing: this problem is too hard to get it done in one go with some particularly smart-ass heuristic. Since there is no one smart-ass heuristic that will save the day, I'd do this:
1) calculate the total trillions to display. (it looks like about 2.5 is the total in the example image)
2) guess a SCPT value to begin with. what about for example "18" based on the actual image at the screen size we see above as posted inside your question.
3) put the big one (sun) in the dead center, and for the middle ones (planets) -- just choose a very easy heuristic, what about from biggest to smallest going anticlockwise srtaing at the top left (don't try to get clever than that with that part of the problem - which indeed could be a huge research project purely on it's own) .. and do the same with the small ones (moons).
4) for the sticks between planets and moons - adopt a trivial solution (like "always 0.5 cm"!!) and that's that. with AI you gotta cut your losses .. everywhere! :) Fix the moons to the planets and forget about them.
5) Now a hard part .. run some sort of heuristic over them that evenly balances what you have so far. treat color as mass and no color as no mass and move the "sun" until it is balancedish. (to be clear, as an example that would be likely downwards if you followed the "planet" layout mentioned in 3.) maybe also move all the planet/moon systems in-out to try to balance it.
6) next the iteration. look at that result and decide if you like it! go back to (2) and pick a new value. (maybe "16!" for example)
(7) there are two possible outcomes here. it might be that during development, there is one magic value for SCPT that always works. perhaps "14.3" or "18.2" or whatever. if you find such a value, never tell anyone. keep it as your own secret information!!!! milk it for everything it is worth with clients. conversely and more difficultly, you might find you need a different value each time. in that case: your AI will have to on it's own iterate through values until it finds one it likes. (for example, by determining whether all your labels fit or not .. and obvious things like "are they touching" "all on screen" etc.)
Anyway FWIW (perhaps nothing) that is what I would do - an iterative approach based on a first guess for the SCPT.
Incidentally: you may well want to buy and study the classic and brilliant book on this sort of display of information!!! Everyone should have a copy.
Tufte's The visual display of Quantatative Information
by Edward R. Tufte
ISBN 0961392142
Regarding the mechanics of laying out the image. You should use quartz or any other low-level drawing - forget about UIViews and the like. You should surely completely separate the logic from the drawing layer, so (even if you do want to change to UIViews, OpenGLES, or whatever) it's only changing a few lines of code.
Hope it helps somehow.
Notes...
** SCPT .. square centimeters per trillion
Followup...
"To keep the logic separate would you use a manager-type pattern?"
To be honest: if I was doing it, I would just start a whole new app purely for the "research" of getting this part, this challenge, working right. In that app (to be honest!) I would make bugger all effort to do anything in any tidy manner whatsoever! :-/ Globals everywhere! :) Unfortunately for me I can only think of the one thing at a time, so at that stage I would only be thinking about the algorithm, per se.
I believe, once you cracked the problem per se, once you came to implement it in a bigger project ... really, FWIW, if it was me, I'd simply make it a class (let's say AmazingClass) nothing more complicated than that. Personally I would set the data somewhere separately (whether in a DB or just an array or whatever) and I would just let the AmazingClass take care of getting the data, even. (My thinking - you never know how the hell you're going to need the data and when, at what point in the process of AmazingClass. So, just give up and let AmazingClass take it as and when it wants it.)
If you are familiar with these awesome-sounding manager-patterns of which you speak - yeah, why not! In short I would heavily separate it out as much as possible. I'm not good enough to speak on the best way to do that - but just completely separate it out somewhere. Sorry I can't help on that one.