• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.21 no.3
• /
• pp.59-66
• /
• 2021

Humans mainly recognize surrounding objects using visual and auditory information among the five senses (sight, hearing, smell, touch, taste). Major research related to the latest object recognition mainly focuses on analysis using image sensor information. In this paper, after emitting various chirp audio signals into the observation space, collecting echoes through a 2-channel receiving sensor, converting them into spectral images, an object recognition experiment in 3D space was conducted using an image learning algorithm based on deep learning. Through this experiment, the experiment was conducted in a situation where there is noise and echo generated in a general indoor environment, not in the ideal condition of an anechoic room, and the object recognition through echo was able to estimate the position of the object with 83% accuracy. In addition, it was possible to obtain visual information through sound through learning of 3D sound by mapping the inference result to the observation space and the 3D sound spatial signal and outputting it as sound. This means that the use of various echo information along with image information is required for object recognition research, and it is thought that this technology can be used for augmented reality through 3D sound.

• Journal of Broadcast Engineering
• /
• v.25 no.5
• /
• pp.750-757
• /
• 2020

One of the basic tasks for robots to interact with humans is to quickly and accurately grasp human behavior. Therefore, it is necessary to increase the accuracy of human pose recognition when the robot is estimating the human pose and to recognize it as quickly as possible. However, when the human pose is estimated using deep learning, which is a representative method of artificial intelligence technology, recognition accuracy and speed are not satisfied at the same time. Therefore, it is common to select one of a top-down method that has high inference accuracy or a bottom-up method that has high processing speed. In this paper, we propose two methods that complement the disadvantages while including both the advantages of the two methods mentioned above. The first is to perform parallel inference on the server using multi GPU, and the second is to mix bottom-up and One-class Classification. As a result of the experiment, both of the methods presented in this paper showed improvement in speed. If these two methods are applied to the entertainment robot, it is expected that a highly reliable interaction with the audience can be performed.

• Journal of the Korea Society of Computer and Information
• /
• v.27 no.10
• /
• pp.29-41
• /
• 2022

Recently, with the development of deep learning technology, a variety of huge models with excellent performance have been devised by pre-training massive amounts of text data. However, in order for such a model to be applied to real-life services, the inference speed must be fast and the amount of computation must be low, so the technology for model compression is attracting attention. Knowledge distillation, a representative model compression, is attracting attention as it can be used in a variety of ways as a method of transferring the knowledge already learned by the teacher model to a relatively small-sized student model. However, knowledge distillation has a limitation in that it is difficult to solve problems with low similarity to previously learned data because only knowledge necessary for solving a given problem is learned in a teacher model and knowledge distillation to a student model is performed from the same point of view. Therefore, we propose a heterogeneous knowledge distillation method in which the teacher model learns a higher-level concept rather than the knowledge required for the task that the student model needs to solve, and the teacher model distills this knowledge to the student model. In addition, through classification experiments on about 18,000 documents, we confirmed that the heterogeneous knowledge distillation method showed superior performance in all aspects of learning efficiency and accuracy compared to the traditional knowledge distillation.

• KIPS Transactions on Software and Data Engineering
• /
• v.12 no.4
• /
• pp.179-188
• /
• 2023

Recently, fake news disguises the form of news content and appears whenever important events occur, causing social confusion. Accordingly, artificial intelligence technology is used as a research to detect fake news. Fake news detection approaches such as automatically recognizing and blocking fake news through natural language processing or detecting social media influencer accounts that spread false information by combining with network causal inference could be implemented through deep learning. However, fake news detection is classified as a difficult problem to solve among many natural language processing fields. Due to the variety of forms and expressions of fake news, the difficulty of feature extraction is high, and there are various limitations, such as that one feature may have different meanings depending on the category to which the news belongs. In this paper, emotional change patterns are presented as an additional identification criterion for detecting fake news. We propose a model with improved performance by applying a convolutional neural network to a fake news data set to perform analysis based on content characteristics and additionally analyze emotional change patterns. Sentimental polarity is calculated for the sentences constituting the news and the result value dependent on the sentence order can be obtained by applying long-term and short-term memory. This is defined as a pattern of emotional change and combined with the content characteristics of news to be used as an independent variable in the proposed model for fake news detection. We train the proposed model and comparison model by deep learning and conduct an experiment using a fake news data set to confirm that emotion change patterns can improve fake news detection performance.

• Convergence Security Journal
• /
• v.22 no.3
• /
• pp.25-32
• /
• 2022

Recently, the image processing industry has been activated as deep learning-based technology is introduced in the image recognition and detection field. With the development of deep learning technology, learning model vulnerabilities for adversarial attacks continue to be reported. However, studies on countermeasures against poisoning attacks that inject malicious data during learning are insufficient. The conventional countermeasure against poisoning attacks has a limitation in that it is necessary to perform a separate detection and removal operation by examining the training data each time. Therefore, in this paper, we propose a technique for reducing the attack success rate by applying modifications to the training data and inference data without a separate detection and removal process for the poison data. The One-shot kill poison attack, a clean label poison attack proposed in previous studies, was used as an attack model. The attack performance was confirmed by dividing it into a general attacker and an intelligent attacker according to the attacker's attack strategy. According to the experimental results, when the proposed defense mechanism is applied, the attack success rate can be reduced by up to 65% compared to the conventional method.

• Phonetics and Speech Sciences
• /
• v.15 no.3
• /
• pp.75-82
• /
• 2023

Automatic speech recognition (ASR) has been revolutionized with deep learning-based approaches, among which self-supervised learning methods have proven to be particularly effective. In this study, we aim to enhance the performance of OpenAI's Whisper model, a multilingual ASR system on the Korean language. Whisper was pretrained on a large corpus (around 680,000 hours) of web speech data and has demonstrated strong recognition performance for major languages. However, it faces challenges in recognizing languages such as Korean, which is not major language while training. We address this issue by fine-tuning the Whisper model with an additional dataset comprising about 1,000 hours of Korean speech. We also compare its performance against a Transformer model that was trained from scratch using the same dataset. Our results indicate that fine-tuning the Whisper model significantly improved its Korean speech recognition capabilities in terms of character error rate (CER). Specifically, the performance improved with increasing model size. However, the Whisper model's performance on English deteriorated post fine-tuning, emphasizing the need for further research to develop robust multilingual models. Our study demonstrates the potential of utilizing a fine-tuned Whisper model for Korean ASR applications. Future work will focus on multilingual recognition and optimization for real-time inference.

• Korean Journal of Remote Sensing
• /
• v.39 no.6_3
• /
• pp.1693-1705
• /
• 2023

Rural areas, which account for about 90% of the country's land area, are increasing in importance and value as a space that performs various public functions. However, facilities that adversely affect residents' lives, such as livestock facilities, factories, and solar panels, are being built indiscriminately near residential areas, damaging the rural environment and landscape and lowering the quality of residents' lives. In order to prevent disorderly development in rural areas and manage rural space in a planned manner, detection and monitoring of hazardous facilities in rural areas is necessary. Data can be acquired through satellite imagery, which can be acquired periodically and provide information on the entire region. Effective detection is possible by utilizing image-based deep learning techniques using convolutional neural networks. Therefore, U-Net model, which shows high performance in semantic segmentation, was used to classify potentially hazardous facilities in rural areas. In this study, KOMPSAT ortho-mosaic optical imagery provided by the Korea Aerospace Research Institute in 2020 with a spatial resolution of 0.7 meters was used, and AI training data for livestock facilities, factories, and solar panels were produced by hand for training and inference. After training with U-Net, pixel accuracy of 0.9739 and mean Intersection over Union (mIoU) of 0.7025 were achieved. The results of this study can be used for monitoring hazardous facilities in rural areas and are expected to be used as basis for rural planning.

• Journal of the Korean Society of Marine Environment & Safety
• /
• v.30 no.2
• /
• pp.147-156
• /
• 2024

The sustainable management and enhancement of marine resources are becoming increasingly important issues worldwide. This study was conducted in response to these challenges, focusing on the development and performance comparison of fish detection and classification models as part of a deep learning-based technique for assessing the effectiveness of marine resource enhancement projects initiated by the Korea Fisheries Resources Agency. The aim was to select the optimal model by training various sizes of YOLOv8-Seg models on a fish image dataset and comparing each performance metric. The dataset used for model construction consisted of 36,749 images and label files of 12 different species of fish, with data diversity enhanced through the application of augmentation techniques during training. When training and validating five different YOLOv8-Seg models under identical conditions, the medium-sized YOLOv8m-Seg model showed high learning efficiency and excellent detection and classification performance, with the shortest training time of 13 h and 12 min, an of 0.933, and an inference speed of 9.6 ms. Considering the balance between each performance metric, this was deemed the most efficient model for meeting real-time processing requirements. The use of such real-time fish detection and classification models could enable effective surveys of marine resource enhancement projects, suggesting the need for ongoing performance improvements and further research.

• Geophysics and Geophysical Exploration
• /
• v.23 no.3
• /
• pp.157-167
• /
• 2020

With the recent advancement of computer hardware and the contribution of open source libraries to facilitate access to artificial intelligence technology, the use of machine learning (ML) and deep learning (DL) technologies in various fields of exploration geophysics has increased. In addition, ML researchers have developed complex algorithms to improve the inference accuracy of various tasks such as image, video, voice, and natural language processing, and now they are expanding their interests into the field of automatic machine learning (AutoML). AutoML can be divided into three areas: feature engineering, architecture search, and hyperparameter search. Among them, this paper focuses on hyperparamter search with Bayesian optimization, and applies it to the problem of facies classification using seismic data and well logs. The effectiveness of the Bayesian optimization technique has been demonstrated using Vincent field data by comparing with the results of the random search technique.

• Journal of KIISE
• /
• v.43 no.7
• /
• pp.795-800
• /
• 2016

In image classification, recent CNNs compete with human performance. However, there are limitations in more general recognition. Herein we deal with indoor images that contain too much information to be directly processed and require information reduction before recognition. To reduce the amount of data processing, typically variational inference or variational Bayesian methods are suggested for object detection. However, these methods suffer from the difficulty of marginalizing over the given space. In this study, we propose an image-text integrated recognition system using active vision based on Spatial Transformer Networks. The system attempts to efficiently sample a partial region of a given image for a given language information. Our experimental results demonstrate a significant improvement over traditional approaches. We also discuss the results of qualitative analysis of sampled images, model characteristics, and its limitations.