Claim Missing Document
Check
Articles

Found 3 Documents
Search
Journal : JOIV : International Journal on Informatics Visualization

Multi Task Deep Learning with Transformer Encoder Decoder for Semantic Segmentation Indah, Komang Ayu Triana; Darma Putra, I Ketut Gede; Sudarma, Made; Hartati, Rukmi Sari
JOIV : International Journal on Informatics Visualization Vol 9, No 1 (2025)
Publisher : Society of Visual Informatics

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.62527/joiv.9.1.1978

Abstract

Visual understanding is one of the core elements of computer vision consisting of image classification, object detection, and segmentation. The system applies a multilayer process to obtain complex image and video understanding using deep learning methods to convert the images to text. Therefore, this study aimed to extract video in the form of frames followed by the application of Transformer and Inception V3 architectures to the image captioning process. The synchronization was based on Multi-task Deep Learning method developed by combining Convolutional Neural Network (CNN) system in the image area, Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM) in the sentence area, Caption Content Network (CCN), and Relational Network Context (RCN). Moreover, Transformer Encoder-Decoder architecture was used in the process of labeling and determining the relationships between objects. The results of the image-to-text conversion process were determined by comparing prospective translated text with one or more references. This was achieved using accuracy and loss validation tables to provide graphical comparisons between the number of epochs and losses. The test results showed that the validation data accuracy was 70.166% while the loss was 22,648% and this showed more epoch iterations led to greater validation accuracy.Keywords— Visual Understanding, Transformer, Encoder, Decoder
Development of a Life Story-Based Digital Counseling Model to Detect Student Depression Using LSTM Jiwa Permana, Agus Aan; Sudarma, Made; Sukarsa, I Made; Hartati, Rukmi Sari
JOIV : International Journal on Informatics Visualization Vol 9, No 2 (2025)
Publisher : Society of Visual Informatics

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.62527/joiv.9.2.2642

Abstract

This research aims to develop an LSTM-based model to help counselors analyze depressive symptoms in students based on their life stories. Depression often occurs among students, which can affect their lives. However, counselling can overcome these mental problems. In order to support the Indonesian government's programs in the field of mental health, concrete steps are needed. One concrete effort is to prevent children from experiencing depression. Depression can be recognized early through a counselling approach. Currently, counselling can be done using digital counselling technology. Therefore, a reliable model is needed to help counsellors. This research used 2,551 tweets about someone's life story from 2,581 datasets. ANN method with LSTM (Long Short-Term Memory) architecture. This counselling is effective in helping individuals resolve psychological and emotional problems, especially depression. The advantage of LSTM is that it can delete data that is no longer relevant. This method effectively processes, predicts, and classifies data based on a certain time sequence. The dataset was taken from Twitter(X) and then validated by experts before being trained with the model. As a result, the model can recognize the depression levels with a test accuracy of 86%. This research has implications in psychology regarding cases of student mental health in realizing the vision of Indonesia in 2045.
Multi-Document Summarization Using Tuna Swarm Optimization and Markov Clustering Widiartha, I Made; Hartati, Rukmi Sari; Wiharta, Dewa Made; Sastra, Nyoman Putra; Astuti, Luh Gede
JOIV : International Journal on Informatics Visualization Vol 9, No 4 (2025)
Publisher : Society of Visual Informatics

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.62527/joiv.9.4.3365

Abstract

The Internet contains a large number of documents from various sources with similar content. The contents of documents that are almost identical will lead to news redundancy, making it difficult for readers to distinguish between factual information and opinions. Multi-document summarization has been designed to enable readers to easily understand the meaning of news documents without needing to read multiple documents. Multi-document summarization aims to extract information from several texts written about the same topic. The resulting summary report enables users to obtain a single piece of information from multiple similar pieces of information sourced from various locations. Various approaches have been used in creating multi-document summaries. Issues regarding accuracy and redundancy are still a significant focus of research. In this paper, a new multi-document summarization model was built using Tuna Swarm Optimization (TSO) and Markov Clustering (MCL) methods. The dataset of this research is Indonesian language news from various online media sources. Based on hyperparameter tuning using training data, the best TSO model performance was obtained at variable values a = 0.7, z = 0.9, and the optimal number of tuna fish > 80. From the research results, it was found that TSO outperformed other swarm intelligence methods. The use of MCL has proven to be effective, as evidenced by the performance results, where TSO achieved an average ROUGE value 7.95% higher when MCL was applied. In this performance test, four standard evaluation metrics of the ROUGE toolkit were used.