I was building an Image Classifier system for an Onychosis detection system with 17 different labels. I’ve used MobileNetV2 160X160 0.75 and even used EON Tuner but the maximum accuracy I was able to get is around 45.2%.
I haven’t used such an amount of labels earlier so would like to have someone to help me out of this?
Indeed the number of classes is pretty high, I just joined your project to understand a bit your dataset.
Would it be possible that you group some of the labels by “higher level category” and then create separate models for subcategories?
What is your target device? This would not be possible on MCU but on Linux based device it could be done.
For example, you take a picture, classify it first by between the higher categories and then classify it again by subcategories.
Not sure it is applicable for your project though.
I guess that would work, lemme do some more exploration into the dataset and then I might be able to separate it into a category. Anyways, my target device is a Raspberry Pi 3 A+ (will be able to shift to 4 B if more memory is needed!).
@louis Have you had a chance to look into the doc I’ve shared earlier? I’m still figuring out how should I develop the model, should I use a single one to classify all of them or should I use an individual one for each?
Awaiting your response.
I honestly have no ideas, your images share a lot of common features and 17 different labels seems a lot to me. So I am not sure it would easily fall first into one category before being able to classify the sub-category…
But I would go that way to avoid the 17 different classes.
Take some time as well to study the output of the confusion matrix and see if you can group the top-category by images that get confused between each other.
@louisMAJOR UPDATE: Listening to your suggestions, I did check the confusion matrix. I then made a new project, collected some new good data with a lot fewer labels and images and made lots of tweakings. Finally, with EON Tuner, I was able to get an accuracy of 87% using Transfer Learning. I’ve tested out the model, and it seems to be working pretty fine.
Onto the deployment stage currently.
Thank you for your help!