Dot number recognition [closed]
Hi, i want to recognize the dot number of the image, just like
the image may be scaled, rotated, but the dot number is unaltered.
I have used the simpleblob detector to extract the small circular blob, showing
as you can see, for some image, the circular blob may be extracted perfectly, while badly for others.
For dot number recognition, i used bag of words object category framework, 'SimpleBlob' detector, 'SIFT' descriptor and SVM classifier are used, but the result is not very good.
So is there any other method suitable for my problem?
Actually it is because of the fact that the blobs are to close together. What you could do is binarize the image and then apply a erosion to the image so that the blobs seperate.
As a matter of fact, i had binarize the image, but it could not always seperate the blobs(for blurring reason), so i think there is no need to extract each dot blob precisely, i just need to recognize the image which pattern it belongs to. Then i used the feature-classifer pattern recognition method, but the result is not very good for simpleblob detector and SIFT descriptor.
And why do the cubes have such a shearing effect? Is it possible to grab pictures on top?
No, that is not possible. Is there any good feature(scale, rotation invariable) for these patterns?
Can you give some insight in the application domain? As to features, its not that rotation and translation that forms a problem but the shearing effect. You should first try to normalize for that camera distortion.