A Style-Transfer-Based Augmentation Framework for Ship Recognition in Infrared Imagery

Abstract

Maritime security relies on reliable ship recognition, where infrared (IR) imagery is crucial for nighttime and adverse weather monitoring. However, the low resolution and scarcity of labeled IR data severely limit recognition accuracy. This study proposes a multi-stage augmentation framework that integrates style transfer, conditional generative modeling, and class-specific augmentation guided by large language models. Experiments under the IR-only setting show significant gains: for VGG16, overall accuracy improves by 1.4 times, from 56.61% to 79.80%, and nighttime IR accuracy improves by 1.6 times, from 44.81% to 71.43%. These results confirm the effectiveness of the proposed framework and its potential for vision tasks constrained by limited and low-quality data.