基于E^2LSH-MKL的视觉语义概念检测

A Visual Semantic Concept Detection Algorithm Based on E^2LSH-MKL

多核学习方法(Multiple kernel learning,MKL)在视觉语义概念检测中有广泛应用,但传统多核学习大都采用线性平稳的核组合方式而无法准确刻画复杂的数据分布.本文将精确欧氏空间位置敏感哈希(Exact Euclidean locality sensitivehashing,E2LSH)算法用于聚类,结合非线性多核组合方法的优势,提出一种非线性非平稳的多核组合方法-E2LSH-MKL.该方法利用Hadamard内积实现对不同核函数的非线性加权,充分利用了不同核函数之间交互得到的信息;同时利用基于E2LSH哈希原理的聚类算法,先将原始图像数据集哈希聚类为若干图像子集,再根据不同核函数对各图像子集的相对贡献大小赋予各自不同的核权重,从而实现多核的非平稳加权以提高学习器性能;最后,把E2LSH-MKL应用于视觉语义概念检测.在Caltech-256和TRECVID2005数据集上的实验结果表明,新方法性能优于现有的几种多核学习方法.

Multiple kernel learning （MKL） methods have a widespread application in visual semantic concept detection. Most canonical MKL approaches employ a linear and stationary kernel combination format which cannot accurately depict complex data distributions. In this paper, we apply exact Euclidean locality sensitive hashing （E2LSH） algorithm to clustering. And by combining the advantages of nonlinear multiple kernel combination methods, we put forward a nonlinear and non-stationary multiple kernel learning method- E2LSH-MKL. In order to make full use of the information generated from the nonlinear interaction of different kernels, this method utilizes Hadamard product to realize nonlinear combination of multiple different kernels. Meanwhile, the method employs E2LSH-based clustering algorithm to group images into sub clusters, then assigns cluster-related kernel weights according to relative contributions of different kernels on each image subset, thereby realizing non-stationary weighting of multiple kernels to improve learning performance; finally, E2LSH-MKL is applied to visual semantic concept detection. Experiment results on datasets of the Caltech- 256 and the TRECVID 2005 show that the proposed method is superior to the state-of-the-art multiple kernel learning methods.