台灣夜市美食辨識之卡路里監控系統

在現今社會下，越來越多文明疾病侵蝕著我們的健康，健康飲食受到許多人的關注。一般人在做自我健康管理飲食記錄時，往往是透過紙本記錄，或是由網際網路上提供的卡路里計算機，點選下拉式表單，選擇自己吃了哪些食物，來做飲食熱量計算，對於使用者來說相當麻煩且操作複雜。近年來隨著智慧型手機迅速發展，出現只需要掃瞄食物包裝條碼就能記錄卡路里的應用軟體。但是在沒有食物包裝條碼的環境下，就無法使用。因為夜市食物通常是沒有卡路里資訊和食物包裝條碼，所以我們選擇夜市食物作為食物辨識目標。食物影像辨識系統對於飲食攝取量的測量和體重控管是相當關鍵的技術。目前台灣沒有一套公開的食物影像資料庫可提供研究人員可以做食物影像辨識研究。本論文提出一套行動裝置上的台灣夜市美食辨識之卡路里監控系統，收集了台灣十大夜市的食物影像，作為食物影像資料庫，其中包含了2,012張靜態影像，202段360度視頻，並提供全台灣第一套公開的夜市美食影像資料庫。在食物辨識方面，我們擷取了 RGB 色彩直方圖特徵、HSV 色彩直方圖特徵、Bag of SIFT 特徵與紋理特徵，使用支持向量機來做訓練分類。卡路里與六大類基本食物成分份量分析是依據影像中參考標籤與食物所占整張影像比例來估算出食物的份量多寡。飲食記錄則是會記錄使用者每日的飲食習慣供使用者作為健康管理和飲食控制的參考。為了測試低解析度的照相手機實作在系統上的可行性，我們實驗用的取像裝置分別使用照相手機與數位相機，資料的拍攝環境分別有夜市現場與實驗室兩種狀況。在將夜市美食分為七大類的情況下，最高的食物辨識率達97.3580%，照相手機與數位相機的影像辨識率相差不到 6%，所以照相手機實作在系統上是可行的。

關鍵字

食物影像資料庫；食物影像辨識；支持向量機

並列摘要

In recent years, a growing number of civilization diseases threats people's health and health issues on food are drawing people’s attention. Most people write dietary records on papers as a kind of health management, or apply web-based calorie calculator by clicking the drop-down lists to choose the kind of food they eat and calculating the corresponding calories. Users often feel troublesome and stop to use it after a while. With the fast development of mobile devices, there appear user-friendly APPs that scan the barcode on food package and retrieve the calorie information. However, not all food packages are labeled with barcodes. The night market food is a typical example. Therefore, the night market food is taken as the subject for food recognition. Automatic food recognition is a key technology for measuring dietary and supplement intake in obesity study and treatment. Currently, there is no open food image dataset available to researchers in food recognition in Taiwan. In this thesis, we propose a Taiwan Night Market Food Recognition and Calorie Monitoring System on mobile devices. 2,012 sample food images and 202 360 ° food videos are collected from ten famous night markets in Taiwan. They are released as the first public dataset of night market food images in Taiwan. For food recognition, we extract various features from food images, including RGB color histogram, HSV color histogram, Bag of SIFT and texture features, then apply support vector machine to train a food classifier. Calories and weight analysis of the six basic food components are calculated based on the ratio of the size of the reference label and the size of the food in an image. Daily food images taken by users are recorded as a reference for health management and diet control. In order to test the feasibility of food recognition on a low-resolution camera phone, food images are collected by using both a camera phone and a digital camera. Moreover, food images are acquired in both the practical night market and the ideal laboratory environment. We categorized the night market food into seven major categories, and the highest food recognition rate of 97.3580%. The difference of the recognition rates using the camera phone and the digital camera is less than 6%, which demonstrates the feasibility of implementation on camera phones.