I don't know if this is the right place to ask this but lets go. I am a beginner in computer vision and I have a project of fruit recognition based on Kaggle's Fruit 360 dataset. I know that CNNs are the obvious choice but I wanted to try some classical approaches (and practice OpenCV in the process :) ). But I don't know what would be good methodological choices. I thought about using Histogram of Oriented Gradients, Global Color Histogram, but I am not sure if they are relevant. Also, would PCA make sense here? I would be thankful for any ideas.
Okay, I assume you are doing this for curiosity, and not towards any particular goal (homework/thesis/work).
My experience with pre CNN techniques has made me understand that feature design is something you intrinsically relate to the task at hand. In your case, fruit recognition, you could take a look at
- Look at shape descriptors (apples, oranges, etc. are spherical or sort of), bananas are cylindrical, watermelons are ellipsoids (?), etc. Zernike moments, and Fourier descriptors could help maybe. HoG does this already I guess.
- Texture (local binary patterns, etc).
I would probably use local colour histograms. I feel that they might better capture the local variations in colours in a patch (red and green in apples, black and yellow for bananas, etc).
As for PCA, I am not sure in what context you are referring to it, PCA on the features, PCA on the image.