The pasta segmentation competition is image processing problem. The problem consists in evolving a detection algorithm capable of separating pasta pixels from non-pasta pixels in pictures containing various kinds of (uncooked) pasta randomly placed on textured backgrounds. The problem is made harder by the varying lighting conditions and the presence, in some of the images, of "pasta noise" (i.e., small pieces of pasta representing alphanumeric characters) which must be labelled as background.
Any evolutionary technique may be used to obtain the desired pasta detection algorithm. Some post processing (whether automatic or manual) of the evolved solutions is also allowed. The output of evolved solutions should be numeric. A threshold will be applied to the output in order to determine the response of the detector when applied to each pixel.
Evolved solution will ranked on the basis of:
- their accuracy on a separate test set (the test set is very similar to the training set in that it was acquired in the same conditions of light and with the same types of pasta and backgrounds). Accuracy will be evaluated by computing the ROC curve for the detector (which in turn is obtained by evaluating sensitivity and specificity for different settings of the threshold) and then calculating the area under the curve.
the extent to which evolution (as opposed to image analysis expertise, manual modifications, etc) has been the key component in achieving the result.Submission of entries:
Submit the evolved detectors as a stand-alone computer program AND source code file(s). The program's command line parameters must include the filename of the input image, the threshold (a floating point number) and the name of the file where the output should be stored.
A short document (in PDF format, ACM style, max 3 pages) should accompany the submission of the evolved detector to explain how this was obtained.
A set of training images is provided below or here as a zip file. (The zip file includes also some hand-segmented "ground-truth" versions of the images in the training set. A more complete training set will be provided shortly.)