Tiny feature detection on High Resolution images

[Hinode]

Hi I just have a question about my application. I have amount of very high resolution images with image size about 4K by 4K pixels (the image size can vary). And my "defeat target" features are about 5 by 5 pixel areas (the sizes also can vary, the shape can vary and the image intensities of the defeat spots can vary). There are less than 10 "defeat target" features per image. Can I use detectron2 to find out the "defeat target" features from the images? Is is efficient to apply "detectron2" to do the job? If so, do you have some suggestions on the config settings? And how many training images/validating images would be sufficient to train the model? Thanks!