-

-

-

Tourism sector has become one of the most potential income for some countries. One of the ways to increase income from tourism sector is to implement information technology so it can attract more tourists to come. The technology that can be implemented is smart tourism. One of the smart tourism implementations for Indonesia tourism, especially for Monas tourism destination is mobile based Visual Question Answering (VQA) application that can provide detailed information about tourism object from mobile phone camera. Focus of this thesis is to produce training model with good detection accuracy. The result of the model training process will be used as model for object detection model that will be used for doing VQA. Dataset that will be used for this research are 600 pictures containing Monas and 25 surrounding objects called class. The methods that will be used for object detection is using YOLO and RetinaNet, where both of these methods will be compared each other by searching the accuracy from evaluation metric from both method. By using original dataset, in YOLO the mean average precision (mAP) score is between 60.77% to 71.99% with 0.1 to 0.9 confidence level threshold range and in RetinaNet the mAP score is between 72.18% to 92.98%. By using augmented dataset, in YOLO the mAP score is between 52.51% to 93.72% and in RetinaNet the mAP score is between 23,8% to 56,19%. The Area Under Curve (AUC) score for original dataset is 0.99 and 0.96 for augmented dataset using YOLO method. Based on the evaluation result, YOLO can detect objects better than RetinaNet and augmented dataset can produce better detection than original dataset.

This paper describes our system, UI, for task A: Sentiment Classification in SemEval-2020 Task 8 Memotion Analysis. We use a common traditional machine learning, which is SVM, by utilizing the combination of text and images features. The data consist text that extracted from memes and the images of memes. We employ n-gram language model for text features and pre-trained model, VGG-16, for image features. After obtaining both features from text and images in form of 2-dimensional arrays, we concatenate and classify the final features using SVM. The experiment results show SVM achieved 35% for its F1 macro, which is 0.132 points or 13.2% above the baseline model.

This research develops end-to-end deep learning- based text-to-speech (TTS) in Indonesian, Javanese, and Sundanese languages. While end-to-end neural TTS, such as Tacotron-2, has made remarkable progress recently, it still suffers from a data scarcity problem for low-resource language domains. Our preliminary study shows that Tacotron-2-based TTS requires a large amount of training data, a minimum of 10 hours of data are required to train the model to synthesize high quality intelligible speech. In this paper, we propose a hierarchical transfer learning strategy to train TTS model for low-resource languages, such as Javanese and Sundanese, by taking advantages of a dissimilar high-resource language of English corpus and a similar intermediate-resource language of Indonesian corpus. We report that the evaluation of synthesized speech using the mean opinion score (MOS) reaches 4.27 for Indonesian (ground truth = 4.36), whereas the MOS for Javanese and Sundanese are respectively 4.08 (ground truth = 4.38) and 3.92 (ground truth = 4.20). The word accuracy (WAcc) evaluations on semantically unpredicted sentences (SUS) reach 98.26%, 95.02%, and 95.43% for Indonesian, Javanese, and Sundanese, respectively. The results indicate that the proposed training strategy can be effectively applied to the target domain of low-resource languages. Using small amounts of training data, less than 1 hour for each language, TTS models can learn fast and achieve adequate performance.

Traditional food knowledge (TFK) is an essential aspect of human life. In terms of sociocultural aspects, TFK is necessary to protect ancestral culture. In terms of health, traditional foods contain better and more natural ingredients compared to the ingredients of processed foods. Considering this background, in this study, data acquisition and automatic food recognition were performed for traditional food in Indonesia. The food images were captured in a professional mini studio. The food image data were captured under the same light intensity, camera settings, and shooting distance from the camera. The parameters were precisely measured and configured with a light intensity meter, adjustable lighting, and a laser distance measurement device. The data of 1644 traditional food images were successfully obtained in the data acquisition process. These images corresponded to 34 types of traditional foods, and 30–50 images were obtained for each type of food. The size of the raw food image data was 53 GB. The data were divided into sets for training, testing, and validation. An automatic recognition system was developed to classify the traditional food of Indonesia. Training was performed using several types of convolutional neural network (CNN) models such as Densenet121, Resnet50, InceptionV3, and Nasnetmobile. The evaluation results indicated that when using a high quality dataset, the automatic recognition system could realize satisfactory area under the receiver operating characteristics (AUROC) and high accuracy, precision, and recall values of more than 0.95.