MS-VSCN:用于影像匹配的多尺度视觉相似度比较网络

MS-VSCN:Multi-scale Vision Similarity Compare Network for Image Matching

影像匹配在多种计算机视觉任务中起着重要的作用。提出一种用于影像匹配的多尺度视觉相似度比较网络。该网络基于孪生网络结构进行构建,将普通卷积与空洞卷积进行融合,使得卷积神经网络在多个尺度上提取到的视觉特征实现互补。首先,网络的输入由两幅待匹配的影像组成,使用权值共享的两个网络分支分别提取两幅影像的深度特征;其次,在网络中与每个普通卷积模块平行设置一个或多个空洞卷积模块,得到的特征经过融合后输入下一个网络层进行特征提取;最后,根据网络输出特征间的欧氏距离对影像的相似度进行比较,从而完成影像匹配关系的判别。实验结果表明,本文提出的方法在测试数据集上的性能优于对比方法,能够有效地完成影像匹配任务。

Image matching plays an important role in various computer vision tasks.A multi-scale visual similarity compare network is proposed for the image matching in the paper,which is based on the Siamese network structure,and combines the normal convolution with the dilated convolution,which makes the visual features of the convolution neural network complementary in multiple scales.The input of the network is composed of two images,and deep features of the two images are respectively extracted by two network branches of shared weight.In the network,one or more dilated convolution modules are set in parallel with each normal convolution module,and the obtained features are input into the next network layer for feature extraction after fusion.Finally,the similarity degree of the image is compared by the Euclidean distance of the network output features to complete the discrimination of image matching relationship.The experimental results show that the method proposed in this paper is better than the comparison method in the test dataset,which can effectively complete the image matching task.