Study Of Feature Extraction Algorithm Based On Metric Learning

Two Dimensional Principal Component Analysis(2DPCA)and Two Dimensional Linear Discriminant Analysis(2DLDA)are two classical matrix-based feature extraction algorithms,which are widely used in image recognition.However,these two algorithms all use the square of the F-norm as a distance metric,the extracted features are greatly affected by noise or outliers,so their robustness is poor.In addition,2DLDA assumes the data obeys the Gaussian distribution globally and only considers the global structure information of data.Therefore,there is a large limitation in practical application.To solve these problems,this paper conducts in-depth research on 2DPCA and 2DLDA respectively and improves the algorithms from different perspectives.The main content of this paper is as follows:1.To solve the problem that 2DPCA algorithm is too sensitive to noise or outliers,this paper used F-norm as a distance metric instead of F-norm square and proposed Angle-2DPCA.This algorithm fully considers the relationship between variance of projected data and reconstruction error.When the image data contains abnormal conditions such as illumination intensity,shooting angle,pose expression,and noise occlusion,the algorithm can obtain a more accurate projection matrix and low-dimensional features.In addition,Angle-2DPCA retains rotation invariance and can better maintain the geometry of data.Experimental results on Extended Yale B,AR,and CMU PIE face databases show that Angle-2DPCA can quickly converge and obtain more robust low-dimensional features with higher efficiency.2.For the problem that 2DLDA does not consider the local structure information of data,this paper proposed 2DLADA.This algorithm uses the weight matrix to represent the relationship between the neighboring samples.It can learning weighted matrix adaptively and pulls the intrinsically similar points close to each other after linear transformation.In addition,considering the robustness of algorithm,this paper used F-norm as a distance metric and proposed 2DLADA-F.Compared with 2DLADA,this algorithm has better robustness and can extract more accurate low-dimensional features.Experiments on artificial database,face databases(Extended Yale B,AR,CMU PIE)and non-face databases(USPS,COIL20)show that both 2DLADA and 2DLADA-F have good performance and can quickly converge.