Help needed: Ideas for BLOB based object classification

  • Hello,


    I am currently working on my master thesis at Silicon Software. I would like to ask you for help!


    The thesis is about BLOB based object classification. The main part is to select the most important parameters like shape factors, etc. and implement their calculation in the FPGA via VA.


    So I would like to ask you two questions:

    1. What do you think are the most important BLOB parameters to use for object classification?
    2. Do you have example images, maybe of field applications, for the usage of the parameters mentioned in 1)?.


    Thanks in advance

    Simon

  • Hi Simon,


    The attached VA Design Pre-Processing_ObjectDetector_v017.va, top-level view:

    pasted-from-clipboard.png


    will work together with image BlobTest.tif.zip nicely and is nearly self explaining within the H-boxes:

    Segmentation + Classification + ObjectSorage


    Concerning your points:


    To 1.:

    A table inside H-box Classification will show where (parallelism + bits) you can find each object feature for classification.

    The features of general interest are normally Area, COGxy, Width and Height. Additional features to the existing ones would be beneficial, but these are not delivered by a single-pass approach. While the other features like Contour-Pixels (orthogonal/diagonal) and the X/Y offset within the binary image are available already. Compacticity can be calculated.

    I am really looking forward to see your additional results.

    From the application specific view a label-image would be the most requested feature, as far as I know.


    To 2.:

    Above you can find a test-image I am currently using as source for implementing an innnovation-time related project.

    I can not share real customer data.


    You can contact me directly for additional questions.

  • Hello,


    in the past weeks I worked out a list of features in binary, gray and color domain. For clarification: they are not implemented in VA so far (except the features obtained by the BLOB operators)!


    I would appreciate further help from experienced people in object classification, there are the following questions:

    1. If you could choose 10 of the listed features, which of these would it be?
    2. Do I miss important Features?

    If you want to help me, please use the attached .pdf, annotate it and post it here or send it via e-mail to me (address on request).


    Thanks in advance

    Simon


    Featurelist_Forum.pdf

  • Hello SWe,


    Sounds like a demanding project:


    As requested *known single pass features* plus add-on features, PDF annotation did not work for me:

    • TOP 10 (ordered by personal impression of importance, high = 1 ... low = 10)
    1. Label Image with correspondance to object feature structure
    2. Hole count
    3. Hole area
    4. Smallest bounding rectangle (width, height, angle, area)
    5. Angle of orientation moment equivalent ellipse
    6. Main and secondary axis of moment of inertia
    7. Value statistics (min, max, mean, median, standard deviation) for each channel in RGB or HSV
    8. Gray statistics (min, max, mean, median, standard deviation)
    9. Radius of min enclosing circle
    10. Radius of max inner circle

    afaik, your PDF is not missing basic features.


    I am curious what your results will be.

  • Dear SWe,

    TOP 10 (ordered by personal impression of importance, high = 1 ... low = 10)
    Label Image with correspondance to object feature structure

    From my point of view a label image related to the corresponding object (found by 1D or 2D Blob-Analysis) would be a drastic benefit to the VA functionality in general, since it would enable additional investigations and feature extractions.

    Please have a look into this powerfull approach and especially into the citations section.

    For further reading; if access given: >>

  • Dear B.Ru,


    we are now able to extract the objects themselves. Which now enables the possibility of calculating features on the input data, like the mean gray value and non additive features of the object.


    Right now I started implementing the gray value statistics, as they are straightforward to implement.


    Best regards

    Simon


    For other readers: I would like to encourage you to help me with the Questions in #3, having more replies than just one would help me much! Thank you :)