1. UKM Learning and Research Repository
2. PHD Thesis / Tesis PHD
3. Faculty of Engineering and Built Environment / Fakulti Kejuruteraan dan Alam Bina

Title:	Automatic food detection and volume estimation using stereovision analysis for dietary and calories assessment
Authors:	Mohammed Ahmed Subhi (P83084)
Supervisor:	Sawal Hamid Md Ali, Assoc. Prof. Dr.
Keywords:	Universiti Kebangsaan Malaysia -- Dissertations Dissertations, Academic -- Malaysia Food Stereovision Dietry Calories
Issue Date:	24-Apr-2020
Description:	Consuming the proper amount and right type of food has been the concern of many people, especially the health-conscious generation. In addition to physical activity and exercises, maintaining a healthy diet is necessary to avoid obesity and other health-related issues, such as diabetes, stroke, and many cardiovascular diseases. Advancements in machine learning applications and technologies facilitated the development of automatic or semi-automatic dietary assessment solutions, a more convenient approach to monitor daily food intake and control eating habits. However, to an extent, existing manual and semi-automatic vision-based approaches require previous knowledge of the consumed food information or an additional image calibration, which is time-consuming, and human intervention or adjustment is still needed. In this research, an automatic food detection, volume, and calories assessment techniques are proposed. The proposed method includes four main phases of dataset acquisition, food detection and classification, food volume estimation and calories assessment, and performance evaluation of the system. A total of 1740 food images were acquired and distributed over eleven food classes. These images are annotated for the purpose of food object detection and classification. A deep convolutional neural network (CNN) was used as the feature extractor and image classifier. Two existing configurations (MobileNets and RFCN) were used to train the object detection and classification model. In the third phase, after the food has been identified and localized, a food volume estimation and calories assessment model is implemented. Based on the acquired volume and food density, the calculated weight is used to estimate the nutritional values, including calories, acquired from online nutritional databases. The last phase is the performance evaluation; in this phase, the performance of the previous research phases is analyzed. Each category of the proposed Malaysian food dataset was evaluated independently to validate each food class. The food detection and classification model was evaluated, and a 5-fold cross-validation scheme was followed, where all the image data were used for both training and testing the model. The volume estimation phase was evaluated by calculating the measurement uncertainty (absolute measurement error) for each of the estimated values (volume, weight, calories). The results show that the proposed food image detection and classification model was able to identify and localize food items in a captured image with an average of 96.2% mean average precision (mAP) and average F-1 Score of 80.5% when training the model with the proposed food image dataset categories individually, while when training the model with 11 combined food categories, the average mAP of the detection and classification model was 94.0% for SSD MobileNets and 92.4% for R-FCN after 80000 iterations for all food categories. For the food volume estimation, the mean measurement uncertainty was only 9.08%, while for food weight estimation the measurement uncertainty was 10%, and it was 8.5% for the calories estimation of four food categories. Hence, the proposed method was able to provide an automatic approach to identify the food category and its relevant nutrient information without user intervention.,Ph.D.
Pages:	143
Publisher:	UKM, Bangi
Appears in Collections:	Faculty of Engineering and Built Environment / Fakulti Kejuruteraan dan Alam Bina

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.