The Study On Feature Descriptor Leaning Based On Deep Leaning

Feature matching is a basic image processing technology in computer science and artificial intelligence,which plays an important role in many applications,such as image retrieval,image registration,and 3D reconstruction.As the computer science and artificial intelligence are widely used in the real life,people begin to pursue the efficient,stable and generalized feature matching methods.During the matching process,it is difficulty to verify that two features belong to the same object,especially when the matching features have a diverse appearance and the non-matching features are similar in original space.Thus,it becomes very important to design high-quality descriptors for local features.Compared with hand-crafted methods,learning-based methods usually have the capability to learn invariant descriptors in complex scenario,so that it becomes the hot topic in current researches.This paper study the descriptor learning methods based deep learning to match the same features efficiently and accurately.The main works in this paper are as follows:1.This paper introduces and analyzes the current descriptor learning methods.According to representation of the output,they are categorized as floating point and binary;based on design process,they can be divided into hand-crafted and leaning-based.This paper discusses the features and theory of different categories,and introduces the representative methods from different categories.2.Propose an descriptor learning method based on network fusion.Inspired by the fusion methods in traditional machine learning,this paper proposes to fuse the networks to learn descriptors.Based on the designed network architecture,this paper generates new samples to train the model by using the data augment.The triplet loss function is employed to training the model in view of its superiority in matching tasks.In addition,to train the model more effectively,the hard negative mining is utilized to obtain the negatives in triplets,further promoting the performance of the proposed method.3.Propose a descriptor learning method based on Auto Encoder.Considering the limited dimension of the descriptors,this paper tries to make the descriptors carry more information about the original images.By constructing the decoder based on deep leaning,the images are reconstructed from the descriptors.Minimizing the reconstruction loss between original images and reconstructed images,the descriptors are forced to carry more information about the original images.The proposed method is evaluated on two standard benchmarks.The experimental results illustrate the superiority of the proposed method.