基于VGG16改进的特征检测器

Improved Feature Detector Based on VGG16

特征检测是计算机视觉中的一个基本问题,特征检测器检测提取特征的能力直接影响着后续图像处理的效果.目前,手工制作的检测器提取可重复性特征效果差,而一些基于学习的检测器网络结构复杂、检测速度较慢.针对这些问题,提出了一种网络轻量化的特征检测器VGG(visual geometry group network)-Det.通过对传统VGG16网络模型进行层间删减以及结构调整,设计了一个紧凑型的网络模型,并使用随机变换的基准图像块作为输入来训练网络,以实现对图像信息的高效率学习和局部特征的快速检测.实验中,对VGG-Det在Webcam、EF(edge foci)、VGG-Affine 3个经典数据集上进行了性能评估,分别取得了67.9%、66.9%、47.1%的平均可重复性.相较于较新的TILDE-P24检测器,所提检测器最高高出11.4%的可重复性,检测速度提升近21.5%,具有明显优势.通过实验结果及其原理分析表明,本文提出的VGG-Det检测器高效可行,同步提升了图像特征检测的速度与可重复性性能.

Feature detection is a basic problem in computer vision.The ability of feature detectors to detect and extract features directly influences the effect of subsequent image processing.Currently available hand-crafted detectors are poor in extracting repeatable features,and some learning-based detectors have complex network structures and slow detection speeds.In this study,a lightweight network feature detector VGG(visual geometry group network)-Det based on deep learning is proposed to deal with these problems.Through the inter-layer reduction and structural adjustment of the traditional VGG16 network model,a compact network model is designed.Randomly transformed standard image patches are used as the input to train the network to learn image information with high efficiency and extract local features rapidly.In this experiment,the performance of VGG-Det on the three classic datasets of Webcam,EF(edge foci),and VGG-Affine is evaluated,and average repeatabilities of 67.9%,66.9%and 47.1%are achieved,respectively.VGG-Det has a repeatability of 11.4%higher than that of the newer TILDE-P24 detector.In terms of time performance,VGG-Det has a significant advantage over TILDE-P24 with an equivalent repeatability.The detection speed is increased by nearly 21.5%.Experimental results and principal analysis show that VGG-Det is efficient and feasible.It synchronously improves the speed and repeatability of image feature detection.