I am working on an image recognition algorithm to classify images of starch granules to their source plant species. My model right now has 10 classes (plant species). Each class is trained with 600 images, then validated with a different 400 images (1000 images/class, 10k images total). I am using Matlab with the pretrained resnet-18 model
Overall, the model is working well. Most plant species are scoring >95% validation accuracy with some minor confusion between closely related species. Two unrelated species are giving me very serious problems, however. Louisiana broomrape has a nearly perfect recall, but its precision is very low. This is mainly from misclassifications of Philadelphia lily. Validation images of Philadelphia lily are, about %75 of the time, being classed as Louisiana broomrape.
I've tried doubling sample sizes (previously, I've run resnet-18 with 300 training and 200 validations images per species) but that hasn't made much of a difference. Philadelphia lily's VA has only improved to around %20 rather than %15
(Similarly, skunk cabbage is misclassifying as Philadelphia lily, but it's much less drastic (~30%). I'm hoping whatever I do to fix the Louisiana broomrape/Philadelphia problem can be applied to fix the skunk cabbage/Philadelphia lily confusion).
Any advice?
