【作者】 龙飞； 林坤辉； 林凡；

【Author】 LONG Fei+, LIN Kun-Hui, LIN Fan (Software School, Xiamen University, Xiamen 361005, China)

【机构】 厦门大学软件学院；

【摘要】 提出了一种联合二维主分量分析(Joint-2DPCA)的图象描述方法,并将其应用到人脸识别。该方法不仅继承了现有2DPCA方法可直接面向图象矩阵进行操作的优点,而且通过两次图象映射的有效集成达到了图象水平、垂直方向维数的联合压缩,弥补了2DPCA方法只能压缩图象水平方向维数的不足,使得特征数目得到大幅度的降低,匹配识别速度也因此得到了大幅提高。更多还原

【Abstract】 So far as a pattern recognition system is concerned, the object governing the design of feature extraction module is generally that extracted features are representative enough and as few as possible. Principal component analysis (PCA) is a well-known data reduction technique widely used in fields of pattern recognition and computer vision. As a kind of subspace analysis method, PCA is vector-based, which means that when we apply it to image recognition problem, an image must be firstly transformed into a high-dimensional vector by concatenating all row or column vectors of the image. In such a case, the evaluation of covariance matrix is usually not accurate because the number of training samples is relatively small, and the computational complexity for following eigenvalue decomposition is high due to the large size of covariance matrix, which causes feature extraction is a very difficult task. Recently, an image projection technique termed 2DPCA was developed for image representation, which treats images as 2D matrices rather than 1D vectors in analysis procedure, as a result, the speed of image feature extraction is improved significantly compared with conventional PCA. Unfortunately, 2DPCA needs to use large numbers of features to represent an image, causing storage requirements are heavy and also features matching process is time-consuming. Against this problem, we discuss in this paper a new framework for image representation --- joint image projection, under which we extend existing 2DPCA to Joint-2DPCA. The main strategy adopted in our method is that two image projections are applied to an image jointly, so the dimensions of extracted feature matrix along both horizontal direction and vertical direction get compressed, and finally the total number of features can be reduced to a great extent. In addition, in each image projection process, we use image-based between-class scatter matrix rather than image-based total scatter matrix as generation matrix of optimal basis vectors, therefore the time of image feature extraction can also be further reduced. The experimental results on ORL face database show that the proposed method outperforms existing 2DPCA method.更多还原