Research On Key Technologies Of Few-Shot Object Detection

In the field of computer vision,many methods of mapping between image data and category,location,text and other labels by using deep neural network have been widely used.In order to achieve satisfactory results,neural network usually needs to do sufficient iterative training on a large amount of data with complete labels,but the acquisition of data and labels is often time-consuming and laborious.However,humans can learn to recognize new objects from a few samples.Inspired by the rapid learning of human beings,the concept of few-shot classification was proposed,and many studies have been devoted to it,but there is still a lack of exploration in the more challenging task of few-shot object detection.The purpose of few-shot object detection is to learn a network which has the ability to identify and locate objects belonging to target domain with only a few images of target domain and corresponding labels including bounding boxes and categories.First of all,based on the previous studies of few-shot classification and a small number of existing studies of few-shot object detection,we offer the exact definition of the few-shot object detection task and provide evaluation criterion for few-shot object detection algorithm.According to the analysis of the traditional object detection framework and the characteristics of few-shot task,this paper tries to design a model on the basis of two-stage object detection framework which is based on region proposal,and combines meta-learning with its training strategies.Secondly,in order to make full use of the labels of few given images,we utilize graph convolutional neural network with edge features to model the complex relationship between the proposed boxes and the support boxes in the second stage.The information transfer between the boxes is completed through the iterative update of the node features and the edge features in the graph,and the categories of the unknown boxes are predicted by using the edge features obtained after the last update of the graph.Finally,for further improvement,this paper introduces attention mechanism to the region proposal phase,and adds the information of support boxes to the classification branch and the box regression branch of region proposal network,which makes the region proposal network can output bounding boxes belonging to current task category space more precisely and reduce the bounding boxes not belonging to current task category space.Obviously,these reliable region proposals are beneficial for following detection procedures and the overall detection performance.In this paper,a series of experiments are conducted on Pascal VOC and MS COCO datasets,which prove the effectiveness of the coarse-to-fine classification network based on graph convolutional network and the region proposal network based on attention mechanism,and prove the superiority of our method by comparing its detection performance with that of other few-shot object detection algorithms.