| 摘要: | 深度學習被應用在辨識人臉以及辨識情緒是近年非常熱門的主題，已經有眾多成功案例顯示即便在不同的開發平台設計出辨識模型，只要搭配適合之輔助軟體都能達到相同的辨識目的；影像辨識是以龐大訓練資料為基礎進而提高辨識率與辨識種類，沒有數量可觀的訓練資料支援，出現錯誤概率是相當高的，萬一在建構的過程中，沒有提供足夠的訓練資源，導致辨識種類不足，該怎麼補救呢?本研究為了解決這個問題，計畫將辨識種類較少之影像情緒辨識結合人體生理數據，匯入神經網路模型計算後，提升辨識率與辨識種類。本研究的第一部分是影像情緒辨識模型，這是一個能辨識喜(Happiness)、怒(Anger)、哀(Sadness)三種情緒的卷積神經網路，以Googlenet作為模型主體；為提升第一部分的辨識率並提高辨識種類，本研究第二部分為收集脈搏與體溫感測器的生理數據，建立生理數據輔助心理數據辨識模型，藉由導入第一模型辨識結果與生理數據後，評估出六種情緒—幸福(Happiness)、憤怒(Anger)、恐懼(Fear)、悲傷(Sadness)、驚訝(Surprise)、厭惡(Disgust)。為使系統智能化、輕巧化，本研究將兩個辨識模型嵌入樹莓派系統，樹莓派透過GPIO連接兩個生理感測器，專用接孔連接樹莓派相機，USB插入加速運算處理元件，將連接完所需硬體的樹莓派控制板結合電池控制模組後，進入第三部分以樹莓派為控制器的機器人，機器人靠著8個伺服馬達與連桿機構產生動作變化，其動作變化是依據生理數據輔助心理數據辨識模型執行結果；第三部分的機器人採用外型為四組連桿的機器狗，藉由辨識結果改變其動作，透露出受測者的心理狀態，本研究對於偵測到的情緒反饋十分重視，因此將機器狗設定為會隨著偵測到的六種情緒辨識結果採取預設動作，做出與主人當下情緒相呼應的動作。
In recent years, the deep learning used to recognize the emotions of faces has become a very popular topic in recent years in particular. There have been many successful cases showing that identification models are designed even in different development platforms, as long as it is matched with suitable software, the same identification purpose can be achieved. The image recognition is based on huge training data to improve the recognition accuracy, without the support of big data, the probability of error is still quite high. In case, in the process of construction, insufficient training resources are provided, resulting in insufficient types of identification, how to remedy it? This research aims to improve this problem, planning to combine the image Emotional Recognition and human physiological data made from a small amount of training data, and import it into the neural network model for calculation to improve the recognition accuracy and recognition types. This is a convolutional neural network which can identify the emotion of “Happiness”, “Anger”, and “Sadness”, choosing Googlenet as the main body of the image recognition model. In order to improve the recognition accuracy of the first part and improve the identification types, the second part of this research is to collect the physiological data of the pulse and body temperature sensors, and establish a physiological data-assisted psychological data recognition model. By importing the recognition results of the first model and the physiological data, assessing six emotions—Happiness, Anger, Fear, Sadness, Surprise, and Disgust. In order to make the system intelligent and lightweight, this research embeds these two identification models into the Raspberry Pi system. The Raspberry Pi is connected to two physiological sensors through GPIO, and dedicated socket is connected to the Raspberry Pi camera, USB plug-in accelerated computing processing element, after connecting the Raspberry Pi control board with the required hardware and the battery module, then is the third part of the robot with the Raspberry Pi as the controller. The robot relies on 8 servo motors and linkages to generate movement changes, and its movement changes are based on the results of the execution of the physiological data-assisted psychological data identification model. The robot in the third part, uses a robot dog with four sets of connecting rod, and changes its movements through the recognition results, revealing the psychological state of the subjects. This research attaches great importance to the detected emotional feedback. The robot dog is set to take preset actions according to the detected six kinds of Emotional Recognition results, and make proper actions that correspond to the current emotions of the subjects. |
