Image classification with small label set

Question

I know that Deep CNN's is all everyone cares about today, and there are a lot of papers about state of the art CNN's for image classification; Alexnet, (Googles) Inception, Microsofts ResNet etc.

However, I do not feel these networks is the right answer to most peoples problems. While most of the big CNN's focuses on classifying any image from a giant pool of labels - I fell most real world problems want to classify an image from a much more limited set of labels.

Say I want to classify images of animals: My input is probably images of animals, but instead of having 10.000 labels I may have somewhere between 50 and 200. I can of course just use one of the large generic networks, but my intuition you should be able to gain some performance (in model size, memory footprint, training time, and/or error rate) by using a smaller network and/or a network optimized for fewer labels.

Is there any research into this? Anything like the ImageNet classification challenge, but with the focus on a much more limited scope?

Answer 1

This recent article presents some benchmarks on image classification on small databases. It is a clear review with many references about the state-of-the-art methods and presents many comparisons in small databases like medical images sets.

One of the key points is that especially for small database a careful hyperparameter optimization is essential to have an objective comparison between models. The authors also present a benchmark to have a consistent comparison between different approaches. This benchmark consists of 5 datasets spanning various domains.

If the link will break in the future, the article is "Image Classification with Small Datasets: Overview and Benchmark" of Lorenzo Brigato et al., 2022. You can find it on arXiv, a database of pre-print scientific articles.