基於行動裝置與雲端運算之食物成分析系統設計實作

Abstract

隨著大眾對健康的重要日趨重視，許多個人健康控管的產品也應運而生，然而對健康帶有直接影響的飲食控管產品基於許多因素並不普遍；因此，本研究目標在設計並架構結合基於行動裝置與雲端運算之食物成分分析系統，以利使用者能迅速得知飲食成分並紀錄飲食狀況。此研究的系統中包含了三個部分－建立真實數據收集的基礎構造、利用感測器融合估計食物尺寸，以及在雲端伺服器上實作食物類別和營養成分預估，在第一部分，透過建立一套食物圖片並記錄其大小的收集工具，並建立一套食物成分分析工具以提供營養師便於分析使用者飲食情況的同時也幫助我們標記和訓練模型；在第二部分，藉由手機中加速度感測器、陀螺儀等感測器來推估相片的標準單位大小和食物厚度估算食物體積；而在第三部分，透過卷積神經網路來訓練食物類別及區域預測模型，再藉此將食物類別、大小及其他食物資訊計算出熱量等資訊，最後將技術實作重新架構在真實資料收集的系統上。透過本研究的方法能夠讓食物尺寸估算更於便利且讓體積預估誤差維持在6.87%到10.18%，在食物圖片辨識的部分透過我們的改善方式讓準確率達到83.75%到87.82%，並且讓結合體積估算和圖片辨識預測的營養成分預估從82.80%提升到86.97%；而此系統運作時間相較於直接在手機上作預測所花費大約27.48秒到55.93秒，透過雲端伺服器傳送並運算再回傳結果只需要花費約7.79秒。

As the public is raising awareness of health, there are more and more personal health management products all over the world. However, diet controlling products which affects health significantly is not popularized due to many reasons. This study aims to construct a system based on mobile-cloud computing to predict the nutrition of food. There are three sections in the system - ground truth data collection infrastructure, food size estimation with sensor fusion, and nutrition predict computing in cloud. In the first section, a collection tool is constructed to collect food data and measure the image size while an analysis tool is built for dietitians to analyze food nutrition and meanwhile collect labelled dataset for training. In the second section, the thickness of food and its volume could be estimated with mobile sensors, such as accelerometer and gyroscope, using our method. In the third section, convolutional neural network is used to predict the food type and its portion in image, and estimating the nutrition and calories of food with predicted type and size. Our method makes food size estimation more convenient and the prediction error remains in 6.87% to 10.18%. The accuracy of food image classification with our improvement is 83.75% to 87.82% and which increases the accuracy of nutrition prediction, which combines the volume estimation and image classification prediction, from 82.80% to 86.97%. Moreover, the system execution time takes via cloud computing including transfer time costs only 7.79s which is much less than doing nutrition prediction on mobile end that costs 27.48s to 55.93s.