基于SuperPoint的轻量级特征点及描述子提取网络

Lightweight Feature Point and Descriptor Extraction Network Based on SuperPoint

图像特征点及描述子提取是SLAM、SFM和3D重建等任务的基础,较好的图像特征点及描述子提取算法会对这些任务的进步产生十分重要的作用.本文聚焦于提取特征点和描述子算法中鲁棒性较高、性能较好的SuperPoint网络,对该网络进行了一定程度的改进.针对其计算量和参数较大的问题,首先将普通卷积改成深度可分离卷积,改变卷积层数和下采样方式,之后改进通道剪枝算法,使其可以应用于深度可分离卷积,对网络进行剪枝.实验结果显示,在轻微损失特征点检测和匹配效果的情况下,将网络参数量压缩为原来网络的15%,运算量压缩为原来网络的5%, FPS提升6.64倍,取得了较好的实验效果.

The extraction of image feature points and descriptors is the foundation of some tasks such as SLAM, SFM,and 3D reconstruction. Preeminent algorithms for image feature point and descriptor extraction play a significant role in processing these tasks. This study accomplishes some improvements in the SuperPoint network with high robustness and good performance in the extraction of feature points and descriptors. Considering the flaws of the heavy calculation burden and massive parameters, the authors first change the ordinary convolution to depthwise separable convolution, the number of layers, and the down-sampling method. Afterward, the channel pruning algorithm is perfected so that it can be applied to depthwise separable convolution and prune the network. Experiments have proved that this study reduces the network parameter number and calculation burden respectively to 15% and 5% those of the original SuperPoint network,and the FPS is increased by 6.64 times under the condition of a slight loss of feature point detection and matching effects.Thus, good experimental results are achieved.