I'm about to build a system that is supposed to evaluate images (900 x 150) like the following and classify it in to one of five categories:
In case you're wondering, they're snapshots of the clustering distribution across the flow cells of a DNA sequencer machine. The total dataset that are currently being labeled is around 30000 images. What would be the most effective way to get a system to categorize these images?
My initial inclination was to fine tune an existing vision model out there (one of the Hugging Face models for example) but I have some doubts as to how appropriate that would be; most of the vision models are trained on photographs of people and things in the world so it seems like trying to fine tune one of these models to recognize images that look like noise just wouldn't work.
If that presumption is correct, would training a CNN based model from scratch be a better way to go? Perhaps a modified version of a CNN? I read an interesting research paper "Image noise types recognition using convolutional neural network with principal components analysis" where the authors incorporated PCA in to their CNN.
Or are there other alternatives?
