小样本目标检测研究综述

A Survey on Recent Advances in Few-Shot Object Detection

数据驱动下的深度学习技术在计算机视觉领域取得重大突破,但模型的高性能严重依赖于大量标注样本的训练.然而在实际场景当中,大规模数据的获取和高质量的标注十分困难,限制了其在特定应用领域的进一步推广.近年来小样本学习在目标检测领域的发展,为解决上述问题提供了新的研究思路.小样本目标检测旨在通过少量标注样本实现对图像中目标的分类和定位.本文从任务和问题、学习策略、检测方法、数据集与实验评估等角度出发,对当前小样本目标检测的研究成果加以梳理和总结.首先,系统性地阐述了小样本目标检测的任务定义及核心问题,并讨论了当前方法采用的学习策略.其次,从工作原理角度出发,将现有检测方法归纳总结为四类,对这四类检测方法的核心思想、特点、优势及存在的不足进行了系统性的阐述,为不同场景下选择不同的方法提供了依据.之后,本文对目前小样本目标检测采用的典型数据集、实验设计及性能评估指标进行了深入分析,进而对四类典型方法在数据集上的实验结果进行概括总结,尤其是对部分典型方法的检测性能进行了系统性对比分析.最后,立足于现有方法的优势和劣势,我们指出当前方法面临的挑战,并对下一阶段小样本目标检测技术未来的发展趋势提出了见解,期望为该领域的后续研究提供参考.

In recent years,with the substantial progress in large data sets and hardware technologies and the tremendous continuous breakthroughs of deep learning models in various fields,several fundamental computer vision tasks based on deep neural network models have gradually matured.Traditional supervised machine learning models must be trained with large-scale labeled data,while visual data in the real world presents a significant long-tail effect.Data-rich categories occupy the majority of the total categories.However,in practical application scenarios,scarce categories may make data acquisition and labeling difficult due to privacy,security,high labeling cost,and other factors.Accessing large-scale data and high-quality annotated samples is often challenging.In few-shot learning scenarios,the traditional deep learning algorithm cannot be fully trained,which makes the deep neural network easy to overfit,and the generalization ability of the model is seriously affected.The recent deep learning techniques cannot meet the needs of scenarios with fewer labeled training samples.Unlike deep neural networks,the visual system of humans can exhibit a remarkable ability to learn novel concepts from a few examples quickly.Such data-efficient ability is precisely what the practical application needs.The universality and generalization capabilities of existing data-driven deep learning technology are far from reaching the level of human cognitive learning.Inspired by the human learning mode,few-shot learning is gradually gaining attention in the academic field.With the deepening of the research,developing few-shot learning in object detection provided a new research idea for solving the above problems.Few-shot object detection aims to classify and locate objects in images by a small number of labeled samples.In the scenario of data scarcity,how to exploit a few labeled samples to learn,design a detection model with good generalization ability,and extend it to new tasks,is an urgent problem to be solved in few-shot object detection.In this paper,we sort out the research findings of few-shot object detection from the perspectives of tasks and problems,learning strategies,detection methods,datasets,and experimental evaluation.First,the task definition and core problem of few-shot object detection are systematically described,and learning strategies are discussed.Second,from the principle perspective,the existing few-shot object detection methods are summarized into four categories,including meta-learning based,transfer-learning based,data augmentation based,and metric-learning based methods.The core ideas,characteristics,advantages,and shortcomings of the four detection methods are systematically elaborated,providing a basis for choosing different methods in different scenarios.After that,this paper provides an in-depth analysis of the typical datasets,experimental design,and performance evaluation indexes currently used for few-shot object detection.Then we summarize the experimental results of four types of typical methods on datasets and systemically compare and analyze the detection performance of some typical methods.Finally,we point out the challenges of the current methods and provide insights on the future development trends of few-shot object detection based on the advantages and disadvantages of the existing methods.It is expected to provide references for subsequent research works in this field.