CAJ | 학술논문

LPP算法是无监督算法，并没有考虑到不同类别的样本对分类效果的影响，结果会造成不同类数据点的重叠，故所获得的子空间对于分类问题来说未必是最优的。提出一种新的基于监督判别局部保持投影（SDLPP）的表情识别算法。利用样本的类别信息重新构造LPP算法中的相似矩阵，然后在目标函数中增加类间散度约束，这样就会在保持样本点局部结构的同时，使不同类的样本点相互远离，从而得到更具有判别性的表情特征。该算法在识别率上比其他方法都有较大提高，通过在JAFFE表情库上的实验验证了其有效性。
LPP산법시무감독산법，병몰유고필도불동유별적양본대분류효과적영향，결과회조성불동류수거점적중첩，고소획득적자공간대우분류문제래설미필시최우적。제출일충신적기우감독판별국부보지투영（SDLPP）적표정식별산법。이용양본적유별신식중신구조LPP산법중적상사구진，연후재목표함수중증가류간산도약속，저양취회재보지양본점국부결구적동시，사불동류적양본점상호원리，종이득도경구유판별성적표정특정。해산법재식별솔상비기타방법도유교대제고，통과재JAFFE표정고상적실험험증료기유효성。
Locality Preserving Projection(LPP)algorithm is unsupervised, which does not take into account the impact on different classes of samples on the classification effect and results in the overlap of the data points for different classes, so the sub-space for classification problems may not be optimal. This paper proposes a new expression recognition algorithm based on Supervised Discriminative Locality Preserving Projection(SDLPP). The algorithm firstly makes use of the classi-fication information of samples to reconstruct the similarity matrix of LPP, and then adds between-class scatter constraint into the objective function. This will make sample points of different classes away from each other when preservers them in local structure, so as to get more discriminative expression feathers. The method improves recognition rate than others, and the results of the experiments on JAFFE database indicate that it is effective.