I am always amazed by those facial recognition app like google picasa. How is it able to find a human face so fast and accurate? How does the program even segment a face pattern out of the noisy background? Pixel by pixel? color?
I don't know much about machine vision, but I am always curious about this type of apps. As for object detection, I am still in the stage of color seperation and then put a cluster of pixels together. This requires more input and efforts in the image acquisition, like better camera, ideal lighting, etc, but many times the results are still not satisfying. What seems so easy task for human eyes is a Mission Impossible. One of the example is a seedling in the background of a brown paper. If it is that easy to find a human face, it should not be hard in this case to detect another object, seedling. Any one has any suggestions what I should explore next?
Thank you for your interest in machine vision. Often times our latest developments in the area of machine vision are presented during the Vision Summit during NI Week in Austin. A summary of last year's talks can be found here. You can find out more about this year's vision summit here.
Unfortunately, I am not aware of how companies like Google perform their face detection. Facial recognition, as opposed to detection, is much more difficult, and has been developed by companies using NI software before. This article describes the application.
Can you describe more of the constraints for your application with the seed? Can you provide a template image to the algorithm beforehand? If so, you can simply use pattern matching to locate the object.
I guess what I am looking for is the feature-based algorithm instead of the pixel-based one we are using now. I don't think LV vision development module has any feature-based functions, like SURF or other object detection functions. I think MatLab has gone way further than NI in the computer vision field.
Thanks for the resources you mentioned. A template matching methods only works for those simple and stable patterns while a seed or an animal is constanly changing, moving. In those cases, pixel-based detection algorithm becomes too limited.