改进的基于深度卷积网的图像匹配算法_雷鸣.pdfVIP

计 算 机 系 统 应 用 2017 年 第 26 卷 第 1 期 ① 改进的基于深度卷积网的图像匹配算法 雷 鸣, 刘传才 (南京理工大学 计算机科学与工程学院, 南京 210094) 摘 要: 鉴于图像匹配中单一特征难以获得理想效果的问题, 提出一种改进的基于深度卷积网的图像匹配算法. 首先对卷积层作展开, 利用 BLAS(Basic Linear Algebra Subprograms)高效地计算矩阵乘法, 从而提高了算法运行 速度; 然后通过基于 POEM(Pattern of Oriented Edge Magnitudes)特征的匹配点筛选方法, 去除部分误匹配点, 增 强了基础矩阵的鲁棒性. 实际图像的实验验证了改进算法的准确性和实时性, 对于重复纹理及旋转图像的匹配 效果显著. 关键词: 图像匹配; 梯度信息; 深度卷积网络; BLAS; POEM 特征 Improved Image Matching Algorithm Based on Deep Convolution Network LEI Ming, LIU Chuan-Cai (School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China) Abstract: In view of the difficulty of obtaining the ideal effect by the single feature in image matching, an improved image matching algorithm based on deep convolution network is proposed. First of all, the algorithm expands the convolution layers, and efficiently computes the matrix multiplication by using the BLAS (Basic Linear Algebra Subprograms) libraries. The algorithm can accelerate the running speed. Then, a screening method of matching points based on the POEM (Pattern of Oriented Edge Magnitudes) feature similarity of feature points is used as well. The method can remove some wrong matching points, make the estimated fundamental matrix more robust and improve the repeating texture and rotational image. The accuracy and instantaneity of the algorithm are proved by the experimental results. Key words: image matching; gradient information; deep convolution network; BLAS; POEM feature 图像匹配是图像分析和处理的重要内容, 是实现 SIFT 和 SURF 这类描述子都具有旋转和光照不变性, 计算机视觉研究的重要环节. 在实际的应用中, 待匹 可在图像中检测出关键点, 但是其描述子构成复杂度 配的图像常常存在视角、亮度、平移、噪声、旋转等