Looking for ideas about a goal of my project

Hello,

I was given the task of creating a project that will work with feature extraction/detection on Corel-1000 dataset. However, the specific goal of the project was not set, so it's up to me to come up with some idea. I was thinking about comparing two images, finding matching features using SURF or ORB algorithm and then generating the probability that these images belong to the same class. I would like to know what do you think about my idea and hear your opinion if it does make any sense at all. Also if you have any other ideas or notes, please share them.

Thank you.

[Image Analyst]

What exactly is the Corel-1000 dataset. What kinds of subjects do the images contain?

I have no idea what you want to do. Do you want to do the simplest thing possible to meet expectations, like loop over all images and just get the mean value of each image? Or do you want to exceed expectations and do something really challenging and spectacular? Or something in between like just classifying the images into dominant colors? Like this is a reddish image, this other one is a greenish image, etc.

[Samuel G.]

It's dataset which consists of 1000 images each belonging to 1 class. There are 10 classes total. You can find it here.

I've got very vague idea either. I was thinking about grabing 2 images from the database, then extracting matching features and just by number of these matching features trying to decide if those images belong to the same class. Everything without the use of ANNs. However, I'm not sure if it's doable just by using algorithms like SURF or ORB for feature detection and matching.

Thank you.

Answer 1

[Image Analyst]

Sure, give it a shot. It probably won't be as accurate as the NN method but you can show some sort of similarity based on what you measure, which may or may not indicate membership in the predefined class -- it depends on what you measure. For example if you have a database of butterflies classified to what country they are found in, then you probably can't get that from image analysis, but you can determine if butterflies belong to certain color classes, like "blue butterflies", "orange butterflies", etc.