基于卷积神经网络的局部图像特征描述符算法

Image Feature Descriptor Algorithm Based on Convolutional Neural Network

为提升基于图像序列三维重建的速度,解决传统局部特征描述符算法提取速度慢的问题,设计了一种基于深度学习的局部特征描述符网络。利用特征描述符网络实现对图像特征点的特征提取,结合本文采用的欧氏距离匹配准则,实现了对不同图像间特征点的匹配。算法对MVS数据集进行了验证,实验结果表明:提出的局部特征描述符算法实现了对图像特征点特征的快速准确的提取、匹配,与传统特征描述符算法相比,特征提取时间缩短了50%以上,特征点的匹配时间缩短了60%以上。相对于本算法中复杂结构的特征描述符网络,结构简单的泛化性更好,可拓展到航天领域的三维重建中。

In order to improve the speed of three-dimensional(3D)reconstruction based on image sequences and solve the problem of slow extraction speed of the traditional local feature descriptor algorithm,a local feature descriptor network based on deep learning is proposed.The feature descriptor network is used to realize the feature extraction of image feature points.By combing the Euclidean distance matching criteria,the feature points of different images are matched.The effectiveness of the method is verified by an experiment on the multi-view environment(MVS)dataset.The experimental results show that the proposed local feature descriptor algorithm achieves the fast and accurate extraction and matching of the image feature point features.Compared with the traditional feature descriptor algorithm,the feature extraction time is reduced by more than 50%,and the feature point matching time is shortened by more than 60%.Compared with the feature descriptor network of the complex structure of the proposed algorithm,the simple structure generalization of the proposed algorithm is better.The algorithm can be extended to the 3D reconstruction in the aerospace field.