Improving Rare-Class Recognition of Marine Plankton with Hard Negative Mining

Joseph L. Walker, Eric C. Orenstein

Abstract: Biological oceanographers are increasingly adopting machine learning techniques to conduct quantitative as-sessments of marine plankton. Most supervised plankton classifiers are trained on labeled image datasets annotated by domain experts under the closed world assumption: all object classes and their priors are the same during both training and deployment. This assumption, however, is hard to satisfy in the actual ocean where data is subject to dataset shift due to shifting populations and from the introduction of object categories not seen during training. Here we present an alternative approach for training and evaluating plank-ton classifiers under the more realistic open world scenario. We specifically address the problems of out-of-distribution detection and dataset shift under the class imbalance setting where downsampling is needed to reliably detect and classify relatively rare target classes. We apply a hard negative mining approach called Background Resampling to perform downsampling and compare it to other strategies. We show that Background Resampling improves detection of novel particle classes while simultaneously providing competitive classification performance under dataset shift.