# Tag Info

16

I have a feeling that http://www.tineye.com/commercial_api may be the solution here. Simply throw the Twitter profile image to Tineye, see if it returns images (and associated URLs) that can clearly be identified (or automatically scored using simple word-count logic) as being related to (or of) that little sack of **. Simples!

11

Since you are able to filter to only those that are clear portrait photos, I'm assuming you have some method of feature generation to transform the raw images into features that are useful for machine learning purposes. If that's true, you could try to train a classification algorithm (there are lots of them: neural networks, etc.) by feeding the algorithm ...

8

A generalized Hough transform is exactly what you want. The difficulty is to do it efficiently, because the space of circles in 3D has six dimensions (three for the center, two to orient the plane, one for the radius). This seems to rule out a direct calculation. One possibility is to sneak up on the result through a sequence of simpler Hough transforms. ...

8

The standard answer to this question is the chi-squared test. The KS test is for unbinned data, not binned data. (If you have the unbinned data, then by all means use a KS-style test, but if you only have the histogram, the KS test is not appropriate.)

8

A major insight into how a neural network can learn to classify something as complex as image data given just examples and correct answers came to me while studying the work of Professor Kunihiko Fukushima on the neocognitrion in the 1980's. Instead of just showing his network a bunch of images, and using back-propagation to let it figure things on it's ...

6

In the spirit of brainstorming, I'll share some ideas you could try: Try Hue more? It looks like Hue gave you a pretty good discriminator between silver and copper/gold, though not between copper and gold, at least in the single example you showed here. Have you examined using the Hue in greater detail, to see whether it might be a viable feature to ...

6

No, or at least I would say not necessarily explicitly. If you have an image formation model (e.g. derived from the physics of the imaging process), you can pose recognition, reconstruction or detection as an inverse problem using parametric or implicit representations of your "pattern" or object of interest without making any probabilistic modeling ...

6

You're looking for the Kolmogorov-Smirnov test. Don't forget to divide the bar heights by the sum of all observations of each histogram. Note that the KS-test is also reporting a difference if e.g. the means of the distributions are shifted relative to one another. If translation of the histogram along the x-axis is not meaningful in your application, you ...

6

One standard approach is to use a restricted Boltzmann machine to do the feature extraction, and then reconsider the RBM as a neural network and finish the training using back-propagation. See, for example, G. E. Hinton, "To Recognize Shapes, First Learn to Generate images," Progress in brain research, vol. 165, pp. 535-547, 2007. This is an example of ...

5

You could use a method like eigenfaces, http://en.wikipedia.org/wiki/Eigenface. The following has a good walk through of the procedure as well as links to different implementations. http://www.pages.drexel.edu/~sis26/Eigenface%20Tutorial.htm From here it is common to use this in a classification approach, train a model and then predict cases. You could ...

5

The Baum-Welch algorithm and the Viterbi algorithm calculate different things. If you know the transition probabilities for the hidden part of your model, and the emission probabilities for the visible outputs of your model, then the Viterbi algorithm gives you the most likely complete sequence of hidden states conditional on both your outputs and your ...

5

The idea is nice, but the main problem with averaging is that it works (i.e. removes noise so that you get the essence) only if you average the very similar representation of an exactly the same entity. To get it clear, imagine that you perform your process on a watermelon -- for a human, the word "watermelon" works for a full watermelon as well as for a ...

5

Two things, for starters. One, definitively do not work in RGB. Your default should be Lab (aka CIE L*a*b*) colorspace. Discard L. From your image it looks like the a coordinate gives you the most information, but you probably should do a principal component analysis on a and b and work along the first (most important) component, just to keep things simple. ...

5

I suspect it depends on the context. In the example you gave, they are contrasting handwritten characters with photographs. Both obviously have covariance structure. However, there are a finite number of letters (26, 52, etc) and the covariance structure of a letter is, by convention, pretty tightly constrained. The number of possible photographs is much ...

4

If you can handle unconstrained imaging conditions you should perhaps look at LFW and PubFig. If you need controlled imaging conditions you should perhaps look at MultiPIE.

4

I prefer a slight change of notation due to the many $n$'s appearing in the original. Let $\alpha$ and $\beta$ designate the images. Let $i$ and $j$ each designate pairs of indexes into the image rows and columns. (Indexing goes from $1$ to $m$ for rows and $1$ to $n$ for columns.) Let $h$ designate a relative index pair (so that its two entries are ...

4

Not an R package, but D. A. Landgrebe from Purdue (author of Signal theory methods in multispectral remote sensing) has sponsored the MultiSpec freeware. Its a rather clunky GUI but gets the job done for most of the common hyperspectral algorithms.

4

You may have heard it said that neural networks are "universal function approximators". In essence, the Cybenko theorem says that for any function mapping reals to reals, you can approximate it with a neural network with sigmoid activation functions. In fact, it turns out that neural networks allow you to compute any function which is computable by a Turing ...

3

The best place to look for free/open source capabilities of this nature is GRASS GIS. The image processing manual is at http://grass.fbk.eu/gdp/imagery/grass4_image_processing.pdf . Because this is constantly undergoing development, it would be worthwhile posting an inquiry on one of the GRASS user lists (found through links on the home page at ...

3

I am afraid there is no; during my little adventure with such data we have just converted it to a data frame form, added some extra attributes made from neighborhoods of pixels and used standard methods. Still, packages ripa and hyperSpec might be useful. For other software, I've got an impression that most of sensible applications are commercial.

3

You need to put on an algorithm detecting which person that picture is referring to. You can build a model based on different portrait pictures of famous personality and use classifiers to ensure that this picture is referring to one of your database picture. You need to use a certain classifier based on different parameters liked to the face, like distance ...

3

If you want to do it yourself, I would recommend using Intel's free and open source OpenCV (CV for computer vision) project. http://opencv.willowgarage.com/ http://oreilly.com/catalog/9780596516130

3

A recent paper that may be worth reading is: Cao, Y. Petzold, L. Accuracy limitations and the measurement of errors in the stochastic simulation of chemically reacting systems, 2006. Although this paper's focus is on comparing stochastic simulation algorithms, essentially the main idea is how to compare two histogram. You can access the pdf from the ...

3

Interesting problem and good work. Try using median colour values rather than mean. This will be more robust against outlier values due to brightness and saturation. Try using just one of the RGB components instead of all three. Choose the component that best distinguishes the colours. You could try plotting histograms of the pixel values (e.g. one of the ...

3

I am also doing some research on object detection with Hough forests, so I have some observations on the method. The split rule is not written in stone. The reason axis-aligned split rules (one coordinate) are preferred is that they are very simple to evaluate. Taking two coordinate differences is also very simple. However, I found that these split ...

3

Different categories of image features come to mind: Color features such as color histograms which could for instance be in RGB or HSV space Other histogram approaches, e.g. histogram of oriented gradients (HOG) Texture features such as Tamura's or Haralick's SIFT and SURF features are popular as well Luckily libraries exist that provide access to many ...

3

You want to measure the signal to noise ratio on each image. This is akin to asking what the error is of a single number: you don't know. What's the error of five? That doesn't make any sense. In this case it might be more interesting to find out the the SNR of the process. Here's how you might go about that: Start with a "perfect" image. That is, an image ...

3

Just a hint, after reading your comment. Each image (face) is a vector. The different faces are considered the dimensions. You are confused about the fact that you don't reduce the number of pixels, but the number of faces. The idea is, imagine you have to keep an album of K faces, each composed of $N$ pixels. Instead of keeping all the $K$ faces, you just ...

2

Following on from my comment above, to get an idea of the broad number of topics that solutions to this question can touch on you just need to consider a simple example where you have images of cars and images of houses and you want to be able to identify existing and new images. Let’s say images are not the same size. Then you will want to change the image ...

2

Samuel, welcome to the website. There's been extensive literature on functional data analysis originating from statistics in the last 15 or so years -- see the classic monograph by Ramsey and Silverman (2005, 2nd ed), and you can find a bunch of others. In this literature, one observation = one image, and this literature doesn't have any second thoughts ...

Only top voted, non community-wiki answers of a minimum length are eligible