Research On Line Feature Description Algorithms Based On Deep Transfer Learning

The local image feature description algorithm is the core algorithm of image matching technology.It is not only widely used in computer vision tasks such as wide baseline matching,image retrieval,image stitching,3D reconstruction,etc.,but also in practical engineering applications such as product quality inspection,fingerprint unlocking,and automatic workpiece detection etc.The local image feature description method describes the local area as a unique feature descriptor.A robust descriptor should not change for matching patches due to changes in viewpoint,illumination,rotation,blur,and noise,etc.,while for non-matching patches,the distinction between different image patches must be maintained.In recent years,with the successful application of deep learning in computer vision and the emergence of large-scale feature point description datasets,research on local feature point descriptors has made breakthrough progress.However,for the study of the line feature description algorithms,due to the lack of large-scale labeled datasets and the difficulty of uniformly expressing lines due to uncertain endpoints and lengths,the research of local image line feature description algorithm has remained on handcrafted stage.Inspired by the line feature descriptor MSLD,this paper represented a line in the image as a mean patch and standard deviation patch calculated from the neighborhood of all points on the straight line,and converted the line feature description into a point-like feature description.Therefore,the deep transfer learning technology was applied to the study of the line feature description algorithm,which solved the problems of insufficient line feature datasets and difficulty in uniform expression.This paper combined a fully convolutional neural network and deep transfer learning to propose two line feature description algorithms based on deep transfer learning.The main research work includes the following three points:(1)Constructed a line feature dataset for training deep neural networks Research on a line feature description algorithm based on deep transfer learning,on the one hand,an existing feature point dataset is required,and on the other hand,a new line feature dataset is also required.To this end,two high-quality line feature datasets,LFpatches and GDLFpatches,were constructed in this paper through mobile phone shooting or Internet download.The dataset included about 3400 pairs of images with different illumination,blur,viewpoint,rotation,JPEG compression,noise and scale changes,and about 138,000 pairs of different mean and standard deviation patches.(2)Line feature description algorithm T-MSLD based on deep transfer learning In order to improve the stability and robustness of line descriptors on a large number of image changes,this paper proposes a line feature description algorithm T-MSLD based on deep transfer learning.Firstly,referring to the deep learning network L2-Net of the feature point descriptors,a seven-layer convolutional neural network was established as a training network of line feature descriptors.Secondly,the network parameters were initialized using the pre-trained L2-Net network weights.Thirdly,the triplet margin loss was selected as the loss function,and the self-built line feature dataset was used to fine-tune and optimize the network parameters.Finally,T-MSLD was used for the feature matching task.The experimental results showed that the T-MSLD proposed in this paper was better than the existing handcrafted line feature descriptors MSLD and IOCD under the evaluation indicator of the total number of recalled positive matches and mean average precision(m AP).(3)Line feature description algorithm based on FCNNS and deep transfer learning In order for the deep convolutional network to learn both local feature changes and global features,to further improve the generalization ability of the model and the quality of line feature descriptors,this paper used Central-Surround to data augmentation the input data of the model,and at the same time,the deep learning network architecture selected a fully convolutional neural network,and transfer learning was used to complete the research of line feature description algorithm based on FCNNS and deep transfer learning.This paper studied three fully convolutional neural network architectures,namely the Central-Surround Siamese network(CS S-Net),the Central-Surround Pseudo-Siamese network(CS PS-Net),Central-Surround 2-Channel network(CS 2-Channel-Net).Experiments showed that all three models were superior to the existing handcrafted descriptors,and had achieved significant performance improvements in blur changes,viewpoint changes,rotation changes,and scale changes.