On 10/22/2013 4:54 PM, Ankit Agrawal wrote:
Hi Jim,What Joe said is correct when you want to label/classify images, since classifying images by trying to find similarity of the test image with the training images on pixel level would not work even if there is some ordinary geometric transform like scaling or rotation or Intensity changes. This is where Feature Detection and Feature Description algorithms mentioned above come into play. I have implemented a few of them in scikit-image as part of this year's GSoC, though some of them are not yet merged. OpenCV's feature2d module is very comprehensive in such algorithms. From your description, it seems you want to label each pixel and hence extracting on a features pixel by pixel basis is a better way to go. Depending on your data, the feature vector used to describe every pixel might depend on values/properties of its neighbouring patch say 5 x 5.It is not clear whether you have a labelled training data. If you do, you will need to describe each pixel(i.e. data point) using a feature vector that is discriminates well between different labels and is similar for pixels belonging to the same label. You will have to use the information about the dataset to figure out what features might be appropriate to describe each pixel. If you don't have a labelled dataset, this turns into clustering problem, where the features you choose for representing a pixel will depend on the motivation for clustering.Another good free resource for Computer Vision ishttp://szeliski.org/Book/Hope this helps. Thanks.
Thanks for the feedback and the reference!Each image, to be classified, may be arranged as a data cube (x,y,f) where x and y are the detector rows and columns while f is the filter used to capture the image. Each image feature has a characteristic range of pixel values for each filter. The "fun" part of this is that there may be some overlap in the range of pixel values across features for some filters.
At the moment, I'm simply constructing mock image data that I know the expected classification solution for as the training set.
Cheers, Ankit Agrawal, Communication and Signal Processing, IIT Bombay. ------------------------------------------------------------------------------ October Webinars: Code for Performance Free Intel webinars can help you accelerate application performance. Explore tips for MPI, OpenMP, advanced profiling, and more. Get the most from the latest Intel processors and coprocessors. See abstracts and register > http://pubads.g.doubleclick.net/gampad/clk?id=60135991&iu=/4140/ostg.clktrk _______________________________________________ Scikit-learn-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
------------------------------------------------------------------------------ October Webinars: Code for Performance Free Intel webinars can help you accelerate application performance. Explore tips for MPI, OpenMP, advanced profiling, and more. Get the most from the latest Intel processors and coprocessors. See abstracts and register > http://pubads.g.doubleclick.net/gampad/clk?id=60135991&iu=/4140/ostg.clktrk
_______________________________________________ Scikit-learn-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
