Tips for training on small image patches?

[ugotsoul]

Hello -- I have high resolution images (4k) that have very small errors, on average 8x8 pixels or less. Our dataset has closely cropped bounding box locations which I have extracted into 30x30 patches. In the problem I am optimizing for, recall is more important than precision. I have two questions:

What models would be effective for this task? I've been tinkering with the stfpm model on the patches, and had used padim on the full images with not great results. I'm assuming patch based models are not useful given I've reduced the training set from full images to patches...

Would segmenting the anomalies into masks, however small (less than 8x8) be worthwhile? I haven't seen too many problems that solve for very small defects in a large image.