| Summary: | 碩士 === 國立臺灣大學 === 資訊網路與多媒體研究所 === 106 === A method of face detection and face recognition based on VGG-Face transfer learning [11] is proposed in this thesis. Model training requires high speed and low cost in industry, and transfer learning fits these two requirements. In related work, deep learning models usually need many training data to train well. In our work, we provide a method to achieve similar performance to other approaches with few data (10 images for each class). Due to the lack of training data, we have to augment dataset, such as transformation, rotation, and random patch. To further reduce inconvenience of system online, we use one-shot learning to lower model training to one time. Our experimental results can work well generally. Accuracy of our transfer learning with small data is similar to large data model. Thus our hardware requirement is relatively lower.
|
|---|
