Sonar Images Classification While Facing Long-Tail and Few-Shot
Wenpei Jiao, Jianlei Zhang
Abstract
This work solves the long-tail and few-shot (LTFS) problems faced concurrently in sonar image classification. Although the popular deep transfer learning (TL) alleviates the few-shot problems, it performs poorly in the tail classes. Moreover, current works involving class rebalancing concepts, e.g., resampling and reweighting, are extensively applied to improve tail class accuracy but reduce head class accuracy. Hence, this article investigates the reason for the nonideal performance of the class rebalancing schemes in TL via an empirical study. Impressively, we discover that while using sonar images, these methods hinder the representation fine-tuning but conditionally promote the classifier fine-tuning. Inspired by our discovery, we introduce a two-stage decoupled training approach for the sonar image classification task and propose a novel multibalanced sampling method. Moreover, based on these two key ideas, we suggest a new pipeline entitled balanced ensemble transfer learning (BETL), which simultaneously overcomes the LTFS problems. Extensive experiments on three sonar image datasets of different sizes and imbalance factors demonstrate that BETL significantly outperforms the existing methods. Moreover, BETL’s effectiveness and portability are verified through several experiments. Our code will be available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/Jorwnpay/TGRS_BETL</uri> .