hi, anyone please tell me.... is feature extraction necessary for image classification??? what if we have an image of size 9X80(pixels), should we do feature extraction???? reply soon thanx.

hi, i have images of size 8X90 and i have to classify these images into five classes, i am using feedfwd backprop NN. can i do it without feature extraction? or should i do feature extraction??? please reply soon. thanx.

[Jan]

Reply soon?! No, I go for a meal at first. It is hard to push a voluntary helper.

As long as the subject claims, that you are working on 9x80 matrices and the body says 8x90, this detail is not important, am I right? But the decision about the method to distinguish the classes will depend on the contents of the images: Do they have different colors, contain different shapes or differ in the contrast only?

Answer 1

[Walter Roberson]

"Feature extraction" can be said to include the use of any information derived from the image, no matter how simple or complex. For example, there are images that could be classified (in context) just by examining a single pixel, so checking that single pixel could be called "feature extraction".

In order to classify an image without using any information within the image itself ("feature extraction"), one would have to either classify randomly or else classify according to "meta-information" such as image file name or JPEG tag (e.g., date taken, focal length, GPS coordinates.)

It is, in some circumstances, possible to classify images based such meta-information. More often it is not.

Whether one must use the formalism of creating "feature vectors" or the like is a different matter than whether one uses "feature extraction" in its general sense. It is possible (and practical) to extract information from an image without using those kind of formal methods.