Seagrass classification using unsupervised curriculum learning (UCL)

Nosheen Abid, Md Kislu Noman, György Kovács, Syed Mohammed Shamsul Islam, Tosin P. Adewumi, Paul Lavery, Faisal Shafait, Marcus Liwicki

Abstract: Highlights•Developed an unsupervised deep learning framework (UCL) for classifying seagrass and background in undersea images.•Imporved sample selection operation to balance classes by selecting equal samples from each cluster for fine-tuning.•Integrated clustering with curriculum learning in UCL, enabling unsupervised deep model training using clustering labels.•UCL learns features from seabed image to classify seagrass without manual feature crafting or data labeling.•UCL achieved state-of-the-art unsupervised learning results on the benchmark DeepSeaGrass dataset.