摘要
为解决传统特征提取过程中过多依赖人工选择和传统DBN网络易忽略局部特征问题,提高人脸识别率,提出一种基于局部三值模式的深度学习人脸识别算法(LTDBN)。该算法首先把归一化的人脸图像均匀分割为多个小块,对每个小块进行LTP运算,然后用统计直方图获得最后图像特征,将其作为DBN的输入数据,利用逐层贪婪学习法对整个网络进行训练识别。该算法在ORL,Yale,Yale-B等公开人脸库的识别率分别达到了98.75%,100%,96.62%,实验结果表明LTDBN算法不仅识别率明显优于其他现有算法,而且也降低了光照、姿态等因素对实验结果的影响。
In order to solve the problems of the heavy dependence on artificial selection during the process of traditional feature extraction and leaving local features out of consideration in traditional DBN network,this paper proposes a face recognition algorithm based on local ternary pattern and deep learning（ LTDBN） to get higher face recognition ratio. This algorithm firstly segments a normalized face image into multiple small parts equally and carries out LTP algorithm for each part. Then the histogram is used to get the final image features. These image features are served as the input data of DBN. The greedy learning algorithm trains and recognizes the whole network level by level. The recognition ratio reaches 98. 75%,100% and 96. 62% respectively in public face databases including ORL,Yale and Yale-B. The experiment results indicate that LTDBN algorithm is markedly superior to other existing algorithms in recognition ratio and mitigates the negative effects of factors such as illumination and posture.
出处
《计算机与现代化》
2018年第2期112-117,共6页
Computer and Modernization
基金
国家自然科学基金资助项目(61305008)
