Active Learning Strategy Using Contrastive Learning and K-Means for Aquatic Invasive Species Recognition

Shaif Chowdhury, Greg Hamerly, Monica McGarrity; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2024, pp. 848-858

Aquatic invasive species like Dreissenid mussels disrupt ecological balance and damage agricultural infrastructure. Machine vision tools can use water samples images for early detection of invasive larvae. Supervised learning techniques require large amounts of labeled data, which is costly to acquire in the case of invasive species. Additionally, invasive species larvae can be rare among aquatic organisms, leading to the problem of data imbalance. Active Learning (AL) reduces labeled data needs by iteratively selecting and labeling the most informative data for model training. In this paper, we propose an innovative active learning strategy for recognition of aquatic invasive larvae with minimal labeled data, while being robust to data imbalance. Our strategy is based on a combination of supervised contrastive training and K-Means clustering. The key idea of our algorithm is to project the data into a smaller, more discriminative representation using contrastive learning, where we can apply clustering to select the most informative samples. We evaluate our algorithm on invasive larvae data and compare with several state-of-the-art AL methods. Traditional AL methods face challenges in generalization, class bias, and low-budget effectiveness. Our method provides an efficient sampling process that is effective in the classimbalanced, low budget setting. Starting with only 100 samples, after 100 additional active learning samples we get 78% balanced accuracy, which is a 27% improvement over random sampling and 22% over core-set.