Improved Training Loss Function Based on Deep Learning for Face Recognition

• Main Authors: Chen, Chun-Yu, 陳俊佑
• Other Authors: Wu, Bing-Fei
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/33tyta

碩士 === 國立交通大學 === 電控工程研究所 === 107 === The difference between face recognition and the other pattern recognition is that the faces of people, which are identified in practical application, don’t belong to the training dataset. To identify these faces, we use the trained deep convolution neural network model to extract the features from faces and then calculate the similarity scores between these features. For enhancing the feature extraction capability of the model when not altering the neural network architecture, we propose the new loss functions, named “Center Line Distance Loss” (CLD-Loss) and “Multiple Center Centralization Loss” (MCC-Loss). Both loss functions are improved the softmax loss. The CLD-Loss makes the extracted feature of the face cluster to the center feature in training period. So the features of same person can be more similar, and the features between different people will be more dissimilar. We combine the CLD-Loss with the Center Loss as MCC-Loss, to let the model learn from the blurred face images for increasing the feature extraction capability of the model. Using these loss function to train the model, the accuracy in the public face recognition test database can raise up to 5%.