Contrastive Learning-based Adversarial Discriminative Domain Adaptation for Underwater Image Classification

Underwater image classification faces the challenge of limited image and label data. Traditional classification methods often struggle to address these issues. To tackle this challenge, we propose a Contrastive Learning-based Adversarial Discriminative Domain Adaptation method. This method transfers easily labeled land images to the underwater domain for classification, employs adversarial learning to reduce the distribution discrepancy between land and underwater images caused by environmental conditions, Triplet samples are constructed using high-confidence pseudo-labels for contrastive learning, enhancing intra-class compactness and inter-class separability, thus optimizing the feature space and improving discriminability. Additionally, samples with similar feature representations but different categories are labeled as hard samples, and different weights are assigned based on the degree of confusion, accelerating the optimization of the feature space. Experimental results demonstrate that this method effectively addresses the underwater image cross-domain transfer classification task, providing new insights for handling limited target domain data and cross-domain distribution discrepancies.