摘要
为应对传统基于图像的火星场景分类算法面对火星复杂未知环境、未见类别频发的情况时表现不佳的问题,研究了多模态信息交互的零样本分类算法,并探索了该算法向深空探测领域的迁移及优化。研究内容主要分为数据集构建和算法研究2个方面。在数据集构建方面,整合并重构了火星探测零样本分类数据集。在算法研究方面,提出了一种基于多模态特征交互的零样本场景分类算法,然后结合知识蒸馏技术,对其进行了模型压缩优化,使其可在保证零样本分类性能的同时,大幅降低模型的参数量和计算复杂度。为验证算法的有效性,进行了大量对比实验和可视化分析,结果证明了所提算法的可行性和有效性。
To address the issue that traditional Mars image classification methods based on images do not perform well in the face of the complex and unknown environment of Mars and the frequent occurrence of unseen categories,a zero-shot classification algorithm of multi-modal information interaction was studied and the migration and optimization of the algorithm to the field of deep space exploration was explored.The research content was mainly divided into two aspects:dataset construction and algorithm research.In terms of dataset construction,the zero-shot classification dataset for Mars exploration was integrated and reconstructed.For algorithm research,a zero-shot scene classification algorithm based on multi-modal feature interaction was proposed.Then combined with knowledge distillation,the model was compressed and optimized,which greatly reduced the number of parameters and complexity of the model while ensuring the performance of zero-shot classification.To verify the effectiveness of the algorithm,a large number of comparative experiments and visualization analysis are carried out,and the results prove the feasibility and effectiveness of the proposed algorithm.
作者
檀晓萌
席博博
薛长斌
李云松
TAN Xiaomeng;XI Bobo;XUE Changbin;LI Yunsong(National Space Science Center,Key Laboratory of Electronic Information Technology for Complex Aerospace Systems,Beijing 100190,China;University of Chinese Academy of Sciences,Beijing 100190,China;School of Cyber Engineering,Xi’dian University,Xi’an 710071,China)
出处
《通信学报》
北大核心
2025年第5期103-114,共12页
Journal on Communications
基金
国家自然科学基金资助项目(No.62401434)。
关键词
火星图像分类
多模态信息
零样本分类
知识蒸馏
Mars image classification
multi-modal information
zero-shot classification
knowledge distillation
