随机亮度差量化的二进制特征描述

Novel binary feature from intensity difference quantization

目的传统的基于浮点型向量表示的图像局部特征描述子(如SIFT、SURF等)已经成为计算机视觉研究和应用领域的重要工具,然而传统的高维特征向量在基于内容的大规模视觉检索应用中存在着维度灾难的问题,这使得传统浮点型视觉特征在大规模多媒体数据应用中面临严峻挑战。为了解决浮点型特征的计算复杂度高以及存储空间开销大的问题,越来越多的计算机视觉研究团队开始关注和研究基于二进制表达的局部特征并取得了重要进展。方法首先介绍了二进制特征的相关工作,并对这些方法进行了分类研究,在此基础上提出了基于亮度差量化的特征描述算法。有别于传统二进制特征描述算法,本文算法首先对图像局部进行随机像素点对采样,并计算像素点对之间的亮度差,通过对亮度差值作二进制量化得到图像的局部二进制特征。结果本文算法在公共数据集上与目前主流的几种二进制特征提取算法进行了比较评价,实验结果表明,本文二进制特征在特征匹配准确率和召回率上超过目前主流的几种二进制描述子,并且同样具有极高的计算速度和存储效率。结论通过实验结果验证,本文二进制特征在图像条件发生变化时仍然能保持一定的鲁棒性。

Objective With the explosive growth of multimedia data on the internet, the efficient organization and retrieval of large-scale image and video data has become an urgent problem, which expects more efficient low-level feature with low computation. This brings a huge challenge to the conventional visual feature. It is urgent to make descriptor more compact and faster and meanwhile remain robust to many different kinds of image transformation. To this end, we first introduce sev- eral schemes of binary features, and then propose a novel fast descriptor for local image patches. Method A string of binary bits is used, which are derived from the intensity difference quantization between pixel pairs that are sampled according to a fixed random sample pattern. Different with the other binary descriptor approaches, our method first extracts the pixel pairs randomly, and then calculates the intensity differences from these point pairs. We quantize these intensity differences into binary vectors to form the local binary descriptor. Result Our experiments show that our method is very fast to extract and it shows better more robustness than the other binary feature schemes. Conclusion The binary descriptor proposed in this paper is computed very fast and it outperform other binary features on the public datasets we used. It proved that quantization-based method can obtain more robustness than compare-based methods.