Classiface: 基於多工卷積神經網路之即時人臉辨識

Classiface: Real-time Face Recognition Based on Multi-Task Convolution Neural Network

人臉辨識技術於近年來快速發展且逐漸受到重視。若以傳統的特徵擷取方法達到人臉辨識，例如：局部二值模式，需蒐集大量訓練資料，且其訓練模型大小容易受資料量影響而變大，使該人臉辨識系統載入電腦時間延長，影響使用體驗。近年來，深度學習演算法的突破，漸漸發展出基於深度學習的人臉辨識，因此本研究提出「Classiface」人臉辨識方法，其主要包含資料預處理、人臉偵測、人臉辨識。預處理方面可細分為訓練與測試神經網路所需資料收集及數據能強方法，我們透過攝影機錄影的方式蒐集大量人臉資料，再以人臉偵測方法擷取影片中每一幀影像之人臉，並使用數據增強方法，擴增資料量及使訓練資料多樣化，提升神經網路的泛化性；人臉偵測部分，我們採用快速且錯誤率低的 MTCNN 方法；辨識方面，透過卷積神經網路，自動提取訓練資料特徵，並建構全連接神經網路，實現人臉分類器。最後，我們以3人且每人 3000 張人臉資料訓練神經網路，達到98%以上的辨識準確率，並能以 Real-time 進行人臉辨識。

Facial recognition technology has received extensive attention and massively developed in recent years. Facial recognizing with a traditional feature extractor, such as Local Binary Pattern (LBP), leads to a strong demand for massive training data. As data increase, the size of training model become even bigger which might more easily cause the drop on the effect of loading system. Recently, the breakthrough of face recognition application development may refer to the evolution of deep learning algorithms. Therefore, we propose the “Classiface” face recognition that includes data processing, face detection, and recognition. We collect the training data straightly from videos so that we can process the samples more efficiently in less time. We also use data augmentation technology to increase the number of samples and to make CNN even more robust. We adopt MTCNN which detects human faces fast and accurately. We also utilize the CNN to automatically extract features from images and use fully connected neural network for face classifying. At last, the experimental results show that our proposed method can achieve average 98% accuracy of face classification in real-time without training more than 3000 samples.