A Research On Few-shot Learning For Remote Sensing Image Retrieval

Remote sensing image retrieval is an important area of image retrieval.It can be used for landmark retrieval,environmental monitoring,urban planning,disaster management and so on.In order to make full use of remote sensing image,it is urgent to use effective information management,mining and interpretation methods.Since AlexNet got the champion of the ImageNet Image Classification Competition in 2012,the Convolutional Neural Network(CNN)has gained its reputation and been widely used in the computer vision field.The performance of content-based remote sensing image retrieval is greatly improved by deep learning technologies especially CNN as well.And Some representative works were produced like R-MAC,VLAD-CNN,Neural Codes and Deep Retrieval.With the limitation of high cost for labeled data,some new research tasks are gradually proposed.For solving the problem,few-shot learning in computer vision provides its theoretical method.What's more,there is a large number of unlabeled data in the realistic application of image retrieval.Therefore,this thesis is to research the few-shot learning in image retrieval.The research goal of this thesis is: how to make full use of the existing historical remote sensing images in the database,only with the help of a small amount of labeled information in the new category images,so as to add new retrieval categories under a small number of manual labels,to achieve retrieval effectiveness for new categories and keep the retrieval effectiveness for the origin categories as far as possible.At the same time,we also need to consider the calculation cost of the model to avoid the extra cost of model training and image feature extraction.At present,existing few-shot learning methods in image retrieval are based on classification tasks,of which representative methods are matching networks,modelagnostic meta-learning(MAML).However,there is little research of few-shot learning for image retrieval in special.Take the difference between image retrieval task and image classification task into consideration,there are some problems for taking few-shot learning methods based on image classification directly for image retrieval: 1.the general index map of image retrieval task is not differentiable for neural network;2.in the training procedure,the while dataset need to be retrieved for the new sample and the retrieval results are effected by other samples;3.the new model after few-shot learning and updating need to re-extract features and generates corresponding representation vectors for all images in the database.This thesis is devoted to solving the above problems,and the contributions are as follows:A few-shot retrieval method based on MAML,aims to solve the problem of high annotation cost and model retraining cost in image retrieval.Model-Agnostic MetaLearning(MAML)framework is employed in the method according to the characteristics of few-shot retrieval.MAML algorithm is an algorithm based on meta learning and its original intention is to find a set of initialization parameters sensitive to few-shot learning tasks,so that only a few data points and iterations can quickly complete the adaptation for new tasks.We combine ResNet and GeM to conduct image feature extraction based on MAML and we use mAP optimization algorithm based on bin to solve the problem that optimization target is not differentiable,so that the whole few-shot retrieval model can be trained end-to-end.The main contributions of this thesis are as follows: 1.There is little existing work for few-shot image retrieval,and the limitations on the definition of the problem is great.Based on it,we redefine the few-shot image retrieval problem formally.2.For the first time,we apply MAML on few-shot retrieval problem.By combining ResNet and GeM to extract image feature,and using the optimization algorithm of directly optimizing the retrieval index mAP,we propose a few-shot retrieval method based on MAML.3.The retrieval effectiveness and the efficiency of model training of our method are verified on multiple remote sensing image data sets.