基于归一化单元转动惯量特征的二值商标图像检索

Binary Trademark Image Retrieval Based on Normalized Unit Moment of Inertia Features

由于计算量的限制和不精确的特征提取方法,使得大多数的图像特征提取方法无法有效地解决图像的检索问题.为了解决这一问题,提出一种基于归一化单元转动惯量特征的商标图像检索方法.该方法首先将商标图像划分为若干个单元子图像,然后分别提取各个单元子图像的归一化转动惯量不变特征,最后使用欧式距离度量图像中的归一化单元转动惯量特征矢量之间的相似性.对图像库中2 000多幅商标图像实验表明,归一化单元转动惯量特征具有良好的旋转、平移、尺度以及变形不变性,得到的检索结果能够很好地满足人的视觉感受.

Owing to the limits of computational complexity and imprecise features extraction method, most of the methods can not solve the problem of image retrieval. To solve this problem, the paper presents a new method based on the normalized unit moment of inertia （NUMI）. The features is presented in this paper. Firstly, each image is partitioned into numerous unit sub-images. Then, the NUMI features of individual unit sub-images of each image are extracted, which are formed into a feature vector for describing the shape of the image. Finally, the euclidean distance is employed to measure the similarity between feature vectors of images extracted. The experiments performed on a database containing more than 2000 trademark images show that the NUMI features keep good invariance under rotation, translation, scale, distortion and the retrieved results match human visual perception very well.