Research On Object Matching Based On Few-shot Learning

Object matching is a basic problem in computer vision,its main task is to detect and locate the object in the given images.Object matching technology is widely used in many fields,such as aerospace,national defense construction and health.In real-world scenarios,images are obtained under different photography conditions,these factors pose a great challenge to object matching algorithms.Simultaneously,we usually only have few image data of the objects,these restrictions request the algorithm to extract useful data from small datasets.This paper conducts in-depth research on object matching algorithms based on few-shot learning,the results obtained are as follows:This paper elaborates the meaning of object matching algorithms,make a brief introduction to the basic flows of object matching algorithms,and performs an analysis of the mainstream algorithms.This paper also makes an overview to few-shot learning,and describes the application scenes of the mainstream few-shot learning algorithms.By combining the object matching mission and few-shot learning algorithm,this paper presents two new object matching algorithms based on few-shot learning,and validates the reliability of the proposed algorithms.The first algorithm combines a siamese network and a traditional object matching algorithm.The siamese network is a special model that can differentiate between its inputs,it has been pre-trained on large datasets and can perform well on new data without any training.To solve the problem that traditional algorithms perform poorly on small dataset,the algorithm proposed in this paper introduces a siamese network to calculates the similarity between the object images and the candidate matches,and finds the most accurate result.Then the effectiveness is approved by experiments.The second algorithm uses transfer learning method to build an object matching algorithm based on an object detection model.The transfer learning methods use parameters learnt from large dataset to help to train a new model on small dataset.This algorithm utilizes the parameters from YOLOv2,by adding new layers and changing the loss function,changes the object detection model into an object matching and localization model.This paper also proposes a new algorithm that uses the Spatial Transformer Network to augment data while training,and improve performance on small object while inferencing.Through a series of experiments,the reliability of the purposed algorithm is approved.