第四次工業革命，加速了人工智慧的研發，透過深度學習及卷積神經網路，開啟了Facenet人臉辨識的新紀元，由於 Facenet 模型的輸出是量化數值，因此能利用此數值來比對多張臉孔的差異度。因此，本研究使用Googlenet建模，透過其能夠將人臉框出的功能，以此來加強表情辨識的準確度；另外，由於大多數的臉部辨識系統的體積過大，或使用雲端網路才能夠完成辨識，但本論文在產品的實作上，為了實現在相機終端上就能實現圖像的識別，不需要連接到雲端的功能。因此論文的第一部分，是採用樹莓派作為本系統主要的控制器，再使用Intel的Neural Compute Stick2 做為視覺處理晶片(VPU)，並選用OpenVINO作為電腦視覺推論及神經網路工具包，在這個裝置上建構臉部辨識系統。藉由上述架構可以使用較小的體積，完成相對輕巧並且不需通過雲端功能，即可使用類神經網路辨識人臉的表情。本研究的第二部分，是將上述輕巧型臉部辨識系統嵌入到機器人中，用以改善互動式陪伴機器人與主人之間的互動能力，將此臉部辨識系統當作機器人的主控制器，陪伴機器人便可根據最新的情感狀態與主人互動，用以改善過去陪伴機器人較為亂數或單一的判斷標準。最後，為驗證其可行性，此論文提供了一些模擬結果，也實作一套雛型產品。
Fourth industrial revolution accelerates the research and development of artificial intelligence. Through deep learning and convolutional neural networks, a new era of Facenet has been opened. Since the output of the Facenet model is a quantified value, this value can be used to compare the degree of difference between multiple faces. Therefore, this study adopts Googlenet to establish model. In addition, due to the large size of most facial recognition systems, or use with cloud networks to complete the recognition. In this research, in order to realize the recognition of images on the camera terminal, there is no need to connect to Cloud. Therefore, in the first part of the research, Raspberry Pi is used as the main controller of the system, Intel’s Neural Compute Stick2 is used as the visual processing chip (VPU), and OpenVINO is selected as the computer vision inference and neural network toolkit. Through the above architecture, a relatively lightweight intelligent face identification system can be established, without need for a cloud host using a neural network to identification facial expressions. Furthermore, in order to improve the interaction requirements between the interactive companion robot and the host, the second part of this study embeds above smart face identification system into the robot as its main controller. Such that robot can interact with the host based on the controller's latest emotional status. To verify its feasibility, some simulation results are provided and a prototype is implemented in this research.
