基于孪生神经网络的小样本目标检测综述

A survey of few-shot object detection based on siamese neural networks

目标检测是计算机视觉的基础任务之一,其主要任务是对图像中的目标进行分类和定位。小样本目标检测的目的就是利用极少数的训练样本实现对目标的检测,从而减少繁杂的标注工作,并实现在只有少量样本场景下的应用。现有的小样本目标检测方法主要包括基于孪生神经网络的方法和基于微调的方法,这些方法通过利用现有的包含大量样本的基类数据集和包含少量样本的小样本数据集的训练,使模型实现对小样本类别的分类和定位。重点调研了基于孪生神经网络的双分支小样本目标检测方法,简要介绍了基于微调的小样本目标检测方案,分析了这些方案的优缺点,指出现有的小样本目标检测方案虽不成熟,模型精度有待提升,性能评估方案也有待完善,但却有着十分广阔的应用前景,未来若能通过深入研究解决小样本目标检测的现有问题,其精度必将赶超传统目标检测。

Object detection is one of the basic tasks of computer vision,and its main task is to classify and locate the targets in the image.The purpose of few-shot object detection is to use a very small number of training samples to achieve the detection ability of the objects,so as to reduce the complicated annotation work,and realize the application in the scenarios with only a small number of samples.The existing methods for few-shot object detection mainly include siamese neural network-based methods and fine-tuning-based methods,which enable models to achieve the classification and localization ability of few-shot categories by using the existing base-class datasets containing a large number of samples and few-shot datasets containing a small number of samples.The two-branch few-shot object detection method based on siamese neural networks was focused on,and fine-tuning-based few-shot object detection schemes were briefly introduced.The advantages and disadvantages of these schemes were analyzed.It is pointed out that the existing small-sample target detection scheme is not mature,the precision of the model needs to be improved and the performance evaluation scheme needs to be improved.However,the small-sample target detection scheme has a very broad application prospect,and in the future,the existing problems of few-short object detection will be solved by in-depth research,so that its accuracy can catch up with the traditional target detection.