摘要
深度学习在基于图像检测和识别的问题上取得了显著的成果,然而该方法通常需要大量标记的样本进行大规模的预先训练,因而难以解决样本量不足条件下的场景感知和认知问题。以大型仓储库区无人机巡逻为背景,面对有效数据稀缺等真实的挑战,针对数据独立性高的多种机器智能学习方法进行探索。面向以仓库巡逻为代表的遥感领域目标识别,系统阐述了小样本学习、零样本学习、单样本学习的问题定义、基于有限示例的高质量样本生成模型和零样本条件下基于语义信息的未知目标识别方法,基于当前研究中出现的不足及挑战,对未来的研究方向进行了展望。
Deep learning has achieved remarkable results based on the problem of image detection and identification,but these methods usually require large quantities of pre-training for large quantities of markers,so it is difficult to solve the scene perception and cognitive problems under the condition of the sample size.In the background of UAV patrol in large storage area,facing the real challenges such as the scarcity of effective data,this paper explores a variety of machine intelligent learning methods with high data independence.For the target recognition in remote sensing field represented by warehouse patrol,the problem definitions of small samples learning,zero sample learning and single sample learning,high quality sample generation model based on limited examples and the unknown target recognition method based on semantic information under the condition of zero samples are systematically described.Based on the shortcomings and challenges in the current research,the research direction in the future is prospected.
作者
刘厦
郝亚峰
仇梓峰
胡炎
LIU Sha;HAO Yafeng;QIU Zifeng;HU Yan(The 54th Research Institute of CETC,Shijiazhuang 050081,China;Key Laboratory of Aerospace Information Applications of CETC,Shijiazhuang 050081,China)
出处
《无线电工程》
北大核心
2022年第8期1402-1408,共7页
Radio Engineering
基金
国家自然科学基金企业创新发展联合基金重点支持项目(U20B2064)。
关键词
仓库巡逻
小样本
目标识别
综述
warehouse patrol
few-shot
target recognition
review
