Patch Matching Based On Learning Dense Features At The Resolution Level

Patches refer to pixel blocks near the interesting points,which can reflect the essential features of the image.Patches matching has a widely applied to image stitching,target identification,three-dimensional reconstruction,and augmented reality.It plays an important role in tasks of computer vision.The main goal of this project is to explore the application of deep learning in patches matching,and improve the accuracy of matching results.Early research on patches matching was based on feature engineering,such as SIFT,which was hand-designed by humans.However,with the increasing number of labeled training data and the increase in computer computing resources,the new generation of deep learning descriptors has surpassed the hand-designed descriptors.This paper focuses on the deep neural network applied in patches matching,aiming at constructing a suitable network structure for patches and improving the accuracy of patches matching.The network structure proposed in this paper is named DFR-net.The proposed network uses a dense connection at the resolution level and has a single-tower structure that allows the neural network focuses on the relationship between patches.The component of DFR-net,named RDCNet block,is experimentally demonstrated suiting for patches matching.This paper also explores the advantages of applying metric learning.To ensure the experimental effectiveness,the DFR-net was trained on the Brown patch dataset and the HPatches dataset.Experiments shows that the proposed network structure has improved the accuracy of patches matching task.