【作者】 杨亮；

【导师】 王勇；

【作者基本信息】 中南大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 目标检测技术是计算机视觉领域中的一个重要研究方向,其主要任务是在图像或视频中识别出目标的位置和类别,然而,基于深度学习的通用目标检测技术通常需要大量的标注数据来进行训练,这限制了其在一些样本稀缺场景中的应用。因此,小样本目标检测技术应运而生,小样本目标检测旨在通过大量的基类数据和少量的新类数据对模型进行训练,得到一个能够在新类数据上达到较高检测精度的模型。现有小样本目标检测方法并没有充分地探索已有样本(支撑集)与测试样本(查询集)间的潜在关联,难以有效地过滤任务无关特征并实现对候选区域特征的精细化对齐表征。基于上述问题,本文提出了一个基于交叉注意力和可变形卷积的小样本目标检测算法,并设计和开发了一个通用的在线目标检测系统。本文的主要贡献如下:(1)针对查询集中任务无关特征未被有效抑制导致生成的候选区域质量低从而影响检测精度的问题,提出了一个基于交叉注意力的特征过滤区域生成网络,该网络能够有效过滤任务无关特征,获得有利于当前检测任务的相关特征,生成更加准确的候选区域,从而提高检测精度。(2)针对候选区域特征未对齐导致检测性能低的问题,提出了一个基于可变形卷积的特征对齐检测头网络,该网络以特征过滤区域生成网络的输出作为输入,利用支撑集中的目标区域特征来指导候选区域特征的形变,实现了对候选区域特征的精细化对齐表征。(3)针对小样本目标检测中支撑集特征与查询集特征间的关联关系未被充分挖掘的问题,提出了一个基于交叉注意力和可变形卷积的小样本目标检测算法。该算法首先使用特征过滤区域生成网络生成候选区域,然后使用特征对齐检测头网络对齐候选区域特征并输出检测结果。(4)设计并开发了一个跨平台通用性高的在线目标检测系统,该系统集成样本管理、样本标注、摄像头管理以及算法模型管理多种功能,在科研项目的结题中提供了有力支撑,为目标检测算法的落地应用提供了便利。为验证本文所提出算法的有效性和先进性,本文基于Pascal VOC小样本专用数据集进行了系列实验,实验表明,本文算法在3-shot、5-shot及10-shot下整体上能够展现出更优的性能,满足了科研项目的应用需求。此外,在两个军事场景数据集中的应用实验也表明了算法的应用价值,为军事领域的小样本目标检测提供了有力的支撑。图63幅,表20个,参考文献72篇更多还原

【Abstract】 Object detection technology is an important research direction in the field of computer vision.Its main task is to identify the location and category of objects in images or videos.However,deep learning-based general object detection techniques typically require large amounts of annotated data for training,which limits their application in scenarios with scarce samples.Therefore,few-shot object detection technology has emerged.The goal of few-shot object detection is to train a model using a large amount of base class data and a small amount of novel class data,so as to obtain a model that can achieve high detection accuracy on novel class data.Existing few-shot object detection methods have not fully explored the potential correlations between the support set and the query set,making it difficult to effectively filter task-irrelevant features and achieve fine alignment representation of candidate region features.To address these issues,this paper proposes a few-shot object detection algorithm based on cross-attention and deformable convolution,and designs and develops a general online object detection system.The main contributions of this paper are as follows:(1)A feature filtering region proposal network based on crossattention is proposed to effectively filter task-irrelevant features,obtain relevant features for the current detection task,generate more accurate candidate regions,and improve detection accuracy.(2)A feature alignment detection head network based on deformable convolution is proposed to achieve fine alignment representation of candidate region features.(3)A few-shot object detection algorithm based on cross-attention and deformable convolution is proposed to fully exploit the correlation between the support set features and the query set features.(4)A cross-platform,highly versatile online object detection system is designed and developed,which integrates multiple functions such as sample management,sample labeling,camera management,and algorithm model management,providing powerful support for the completion of scientific research projects and facilitating the practical application of object detection algorithms.To verify the effectiveness and advancement of the proposed algorithm,a series of experiments were conducted on the Pascal VOC fewshot dataset.The experimental results show that the proposed algorithm can exhibit better overall performance under 3-shot,5-shot and 10-shot scenarios,meeting the application requirements of scientific research projects.In addition,application experiments on two military scenario datasets also demonstrate the value of the algorithm,providing strong support for few-shot object detection in the military field.更多还原