基于局部熵的图像特征描述方法

发布时间：2019-06-26 21:48

【摘要】：使用传统的特征描述方法SIFT在单一尺度上描述图像特征会丢失一部分重要信息,影响图像的正确匹配结果。为了解决这一问题,本文在多尺度模糊空间内提取特征描述子。信息熵从图像显著性角度估计特征点及其周围的信息,能获得更多的图像关键内容,本文提出了基于局部熵的图像特征描述方法。首先,在高斯差分空间(DOG)内计算特征点的多层SIFT描述子,同时统计特征点在每层尺度上的局部熵,计算特征点在每层的熵值占所有层熵总和的百分比,利用所得百分比与每层描述子做乘积;然后,累加所有层描述子;最后,使用平方根算法得到最终局部熵特征描述子。通过与其他描述子的对比实验结果可知,本文提出的局部图像描述方法在精确-召回率、平均均匀准确度和正确匹配数方面具有强鲁棒性。
[Abstract]:Using the traditional feature description method SIFT to describe the image features on a single scale will lose some important information, which will affect the correct matching results of the image. In order to solve this problem, the feature descriptors are extracted in multi-scale fuzzy space. Information entropy estimates the information of feature points and their surroundings from the point of view of image significance, and can obtain more key contents of image. In this paper, an image feature description method based on local entropy is proposed. Firstly, the multi-layer SIFT descriptors of feature points are calculated in Gaussian difference space (DOG), and the local entropy of feature points on each layer scale is counted, and the entropy value of feature points in each layer is calculated as a percentage of the total entropy of all layers, and the product of the obtained percentage and each layer narrator is made by using the obtained percentage; then, all layer descriptors are accumulated; finally, the final local entropy feature descriptors are obtained by using square root algorithm. The experimental results show that the local image description method proposed in this paper is robust in terms of accuracy-recall, average uniform accuracy and correct matching number.
【作者单位】：吉林大学计算机科学与技术学院;吉林交通职业技术学院电子信息学院;
【基金】：国家自然科学基金项目(61101155) 吉林省发展和改革委员会产业创新专项项目(2016C035)
【分类号】：TP391.41

【相似文献】