基于改进Gabor滤波器的多模态特征融合技术

Multi-modal Features Fusion Technology Based on Improved Gabor Filter

传统的Gabor滤波器具有良好的方向特性和尺度特性,然而传统的Gabor滤波器不能提取图像中弯曲区域的局部信息。文中首先对传统的Gabor滤波器加以改进,使其在具有方向和尺度特性的同时具有良好的曲率响应特性,因而对于图像中弯曲的区域能够提取丰富的边缘特征。图像在不同的Gabor滤波器特征下有不同的表现形式,利用Gabor滤波器丰富的多特征信息,可以形成包含丰富信息的多个模态。然后文中提出一个多模态学习(Multi-modal Learning)框架。在此框架内,样本集合被投影到一个公共的鉴别空间内,在这个空间里,来自不同模态的同类样本相互聚集,异类样本相互散开。文中提出的多模态学习框架能很好地利用Gabor滤波器的多特征信息,Poly U掌纹数据库和AR彩色人脸数据库的实验结果表明了该方法的有效性。

Traditional Gabor filter has good characteristics of direction and scale, but cannot extract the local information of bending area for image. Firstly,improve traditional Gabor filter to make it has good curvature response based on good characteristics of direction and scale. So for the image area can extract the edge of the rich characteristics of bending. After filtering with different characteristics of Gabor filter, images have more abundant characteristic information, and contain abundant information of multiple modes. Then propose a Multi- Modal Learning （MML） framework, within this framework, samples are projected onto a common space. In this common space, samples in same class from multiple modals are close to each other, while samples in different classes from multiple modals are far away from each other. Multi-modal learning framework proposed in this paper can make good use of Gabor filter characteristic information. Experimental results with PolyU palmprint database and AR color data set show the effectiveness of the method in this paper.