基于位置敏感哈希的图像语义检索技术研究

项目名称： 基于位置敏感哈希的图像语义检索技术研究

项目编号： No.61301232

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 无线电电子学、电信技术

项目作者： 彭天强

作者单位： 河南工程学院

项目金额： 24万元

中文摘要： 视觉词典法是图像语义检索的主流算法，采用K-Means聚类产生视觉词典的方法虽然获得了较为优越的性能。但是存在同义性、多义性、计算效率低、不支持动态扩展、相似性度量效果差等缺点，因此还很难满足实际应用所提出的准确性、鲁棒性和速度要求。位置敏感哈希由于在高维向量搜索的优良性能被作为一种新型的视觉语义检索模型。本项目主要研究基于位置敏感哈希的图像语义检索技术，搭建面向真实网络数据集的图像语义检索原型平台，具体包括：位置敏感哈希的拓展与改进，学习式度量、视觉语言模型等视觉语义检索算法。一方面通过位置敏感哈希的拓展和改进，使得视觉词典区分性更强、识别率更高、具有可扩展性；另一方面将视觉语言模型和学习式度量等多种方法整合到随机映射中，提高模型和算法的鲁棒性，使得视觉语义搜索系统在复杂情况下仍然能够保持良好的性能，为图像语义检索技术的整体发展提供新的研究思路和研究方法。

中文关键词： 视觉词典；位置敏感哈希；图像语义检索；空间金字塔；深度学习

英文摘要： BoVW(Bag-of-Visual-Words) is the state-of-the-art algorithm in the semantic-based visual retrieval domain. So far, visual words generated by K-Means and similarity measure based on Euclidean distance have achieved advantageouos performance. However, K-Means clustering algorithm has some deficiencies, such as visual word synonymy, polysemy, high computational cost and not supportive of scalability, the maximum and minimum Euclidean distance between any two points tend to be equal in high-dimensional space, which suppress the performance. Locality Sensitive Hash(LSH) is adopted as a new semantic-based visual retrieval model for its good performance in high-dimensional space retrieval task. This project mainly study on semantic-based image retrieval based on LSH, including: development and improvement of LSH, distance metric learning, and visual language model,etc. On one hand, the distinctness of visual words is enhanced and recognition rate and scalability are improved by LSH; On the other hand, the robustness of the model and algorithm is boosted by integrating spatial structure similarity matching and distance metric learning into LSH, resulting in good performance in complicated situation. Research of above technologies brings new vigor into the semantic-based visual retrieval development.

英文关键词： visual dictionary；locality sensitive hash；image semantic retrieval；space pyramid；deep learning