Leveraging the feature distribution calibration and data augmentation for few-shot classification in fish counting

Abstract

Fish counting is critical to the success of fish farming since it serves as the foundation for evaluating fish health and growth, rationalizing feed and water quality management, and so on. To address the issue of unbalance distribution and a limited number of samples in the actual fish counting process, a distribution calibration method based on linear fitting was proposed, achieving accurate classification of few-shot or even no-shot classes without complicated generative models and parameter settings. Firstly, seven features were extracted from the adherent fish, followed by a linear fit of the features to the number of adherent fish; subsequently, the number of adherent fish is considered a class, and the mean and variance of the few-shot classes are calibrated based on the fitted linear relationship; following that, adequate training data for the classifiers are generated by extracting samples from the calibration distribution, and the machine learning classifiers model are trained to accurately classify the number of adherent fish. By testing and analyzing the datasets with 4 sets of long-tailed distributions and nine machine learning classifiers, the P, R, F1 scores of Support Vector Classification(SVC) classifier with Radial Basis Function(RBF) achieved 85.6% (±0.4%), 85.5% (±0.6%), and 85.5% (±0.5%), respectively, while R2 was 0.924 (±0.003), with an average counting accuracy of over 96.2% for each class. In comparison to other classifiers, the SVC classifier demonstrated superior precision and stability.