Have you any idea about implementing 2D object recognition with MATLAB? Which characteristics of objects can feed a neural network?
Asked
Active
Viewed 9,877 times
1
-
1this is a very broad question, check out this question to get started on the subject: http://stackoverflow.com/questions/432569/where-do-i-start-learning-about-image-processing-and-object-recognition – Amro Jul 25 '11 at 14:35
-
Thank you for your comment Armo, but I'm not a totally beginner, I've some backgrounds about both computer vision and neural network, but I never implemented such task. – Maysam Jul 25 '11 at 14:41
-
3Easiest approach training a neural network to recognize objects involves feeding the network labeled images (flattened pixel list) of the object to recognize, along with non-object images. Although you usually get better results if process the raw features: sometimes converting to different colorspace might help, or you can compute the gradient of the image and feed it to the network instead. You can also apply a dimensionality reduction technique like PCA (think eigenfaces)... – Amro Jul 25 '11 at 14:49
1 Answers
3
Object Recognition is generally one of the main part of all image processing task. From Simple cases, like fingerprint recognition and Optimal Character Recognition to movement tracking and etc.
Of course there are many different approaches, considering given problem. It can be only based on color (color base object recogniton), say there is a red ball in a green field so only by detecting green color you can identify the ball like here for tracking ball!. The other simple approach is Morphological Operator. Furthermore, one can uses histogram and from distribution of pixels, find all edges of the picture like here. This method is also used to find the baseline of the written text as well.
More advanced methods are based on Machine Learning Approach. Neural network is most known which basically you train your model by bunch of example, find proper weights/values for neurons and finally asking the model to judge about the new example (test). Of course submitting an image to the network is not really wise; Apart from the computational aspects, there is an over fitting issue. So extracting common pattern among pictures is another challenge. Say, some pattern that all characters "A" are following, could be the curve, angles, intensity, FT values and distinguish it from "L" and so on. This part is also called as Dimension Reduction, since you are mapping/composing all picture pixels into several data point. PCA (principle component analysis) and also check the PCA and SVD in matlab. These methods explain variation of data only in some most high variant basis.
The other point of view of Machine Learning which these days is more hot Is in Statistical approach, by looking at the object as the signal and everything else as noise. Gaussian Process and Kernel Methods are placed in this side. Just search these terms beside "object tracking in Google to get thousands academic paper!.
In sum, its really general question; Depends on the problem, solution can be changed. I suggest you to look through https://stats.stackexchange.com/ since your question can fit to those guys expertise.
