计算机科学
混乱
人工智能
自然语言处理
视觉语言
可视化
自然语言
特征(语言学)
鉴定(生物学)
语义学(计算机科学)
计算机视觉
人机交互
动作(物理)
作者
Xiaohui Xu,Yanming Guo,Xueyi Zhang,Wenzhe Liu,Mingrui Lao
出处
期刊:
[Association for Computing Machinery]
日期:2026-11-06
卷期号:: 12717-12726
标识
DOI:10.1145/3767308.3835844
摘要
Fine-Grained Ship Classification (FGSC) is crucial for maritime safety monitoring, traffic management, and anti-smuggling operations. However, due to the insufficiency of explicit target semantics, current mainstream vision-based approaches inevitably encounter the visual confusion problem when deployed in open-environment applications. Recently, Multimodal Large Language Models (MLLMs) have provided a transformative paradigm for enhancing target semantic understanding in FGSC, offering the potential to optimize visual representations through generation and alignment based on target textual knowledge. However, through careful inspection, we reveal that semantic description hallucinations and coarse alignment granularity are two significant challenges to reliably combine MLLMs with FGSC. To break aforementioned limitations, this paper introduces Anti-Confusion Semantic Generation and Alignment (ASGA), an efficient MLLMs-based FGSC approach containing two innovative modules, termed Multi-Agent Collaboration-driven Knowledge Extraction (MCKE) and Schrödinger Coupling-based Multimodal Alignment (SCMA). Specifically, MCKE seeks to automatically generate consistent and discriminative structured attribute priors for ship visual entities through multiple agents, whereas SCMA overcomes the limitations of traditional point-to-point hard matching by computing a global optimal coupling measure with soft alignments. Extensive experiments validate that our proposed ASGA method achieves remarkably superior performance with strong few-shot capacities over current competitive approaches, reaching 96.28% and 88.99% accuracy on two widely-used FGSC benchmarks, respectively.
科研通智能强力驱动
Strongly Powered by AbleSci AI