• The Journal of the Convergence on Culture Technology
• /
• v.7 no.1
• /
• pp.640-645
• /
• 2021

The deep learning based end-to-end TTS system consists of Text2Mel module that generates spectrogram from text, and vocoder module that synthesizes speech signals from spectrogram. Recently, by applying deep learning technology to the TTS system the intelligibility and naturalness of the synthesized speech is as improved as human vocalization. However, it has the disadvantage that the inference speed for synthesizing speech is very slow compared to the conventional method. The inference speed can be improved by applying the non-autoregressive method which can generate speech samples in parallel independent of previously generated samples. In this paper, we introduce FastSpeech, FastSpeech 2, and FastPitch as Text2Mel technology, and Parallel WaveGAN, Multi-band MelGAN, and WaveGlow as vocoder technology applying non-autoregressive method. And we implement them to verify whether it can be processed in real time. Experimental results show that by the obtained RTF all the presented methods are sufficiently capable of real-time processing. And it can be seen that the size of the learned model is about tens to hundreds of megabytes except WaveGlow, and it can be applied to the embedded environment where the memory is limited.

• Journal of the Korean Society for information Management
• /
• v.35 no.2
• /
• pp.63-88
• /
• 2018

This research evaluated differences of classification performance for feature selection methods using LDA topic model and Doc2Vec which is based on word embedding using deep learning, feature corpus sizes and classification algorithms. In addition to find the feature corpus with high performance of classification, an experiment was conducted using feature corpus was composed differently according to the location of the document and by adjusting the size of the feature corpus. Conclusionally, in the experiments using deep learning evaluate training frequency and specifically considered information for context inference. This study constructed biomedical document dataset, Disease-35083 which consisted biomedical scholarly documents provided by PMC and categorized by the disease category. Throughout the study this research verifies which type and size of feature corpus produces the highest performance and, also suggests some feature corpus which carry an extensibility to specific feature by displaying efficiency during the training time. Additionally, this research compares the differences between deep learning and existing method and suggests an appropriate method by classification environment.

• Journal of the Korea Society of Computer and Information
• /
• v.24 no.11
• /
• pp.51-59
• /
• 2019

Due to the exponential growth of access information on the web, the need for predicting web users' next access has increased. Various models such as markov models, deep neural networks, support vector machines, and fuzzy inference models were proposed to handle web access prediction. For deep learning based on neural network models, training time on large-scale web usage data is very huge. To address this problem, deep neural network models are trained on cluster of computers in parallel. In this paper, we investigated impact of several important spark parameters related to data partitions, shuffling, compression, and locality (basic spark parameters) for training Multi-Layer Perceptron model on Spark standalone cluster. Then based on the investigation, we tuned basic spark parameters for training Multi-Layer Perceptron model and used it for tuning Spark when training Multi-Layer Perceptron model for web access prediction. Through experiments, we showed the accuracy of web access prediction based on our proposed web access prediction model. In addition, we also showed performance improvement in training time based on our spark basic parameters tuning for training Multi-Layer Perceptron model over default spark parameters configuration.

• Journal of Korean Society on Water Environment
• /
• v.37 no.4
• /
• pp.275-285
• /
• 2021

Algal bloom is an important issue in maintaining the safety of the drinking water supply system. Fast detection and classification of algae images are essential for the management of algal blooms. Conventional visual identification using a microscope is a labor-intensive and time-consuming method that often requires several hours to several days in order to obtain analysis results from field water samples. In recent decades, various deep learning algorithms have been developed and widely used in object detection studies. YOLO is a state-of-the-art deep learning algorithm. In this study the third version of the YOLO algorithm, namely, YOLOv3, was used to develop an algae image detection model. YOLOv3 is one of the most representative one-stage object detection algorithms with faster inference time, which is an important benefit of YOLO. A total of 1,114 algae images for 30 genera collected by microscope were used to develop the YOLOv3 algae image detection model. The algae images were divided into four groups with five, 10, 20, and 30 genera for training and testing the model. The mean average precision (mAP) was 81, 70, 52, and 41 for data sets with five, 10, 20, and 30 genera, respectively. The precision was higher than 0.8 for all four image groups. These results show the practical applicability of the deep learning algorithm, YOLOv3, for algae image detection.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.19 no.1
• /
• pp.47-55
• /
• 2024

We spend a lot of time in indoor space, and the space has a huge impact on our lives. Interior design plays a significant role to make an indoor space attractive and functional. However, it should consider a lot of complex elements such as color, pattern, and material etc. With the increasing demand for interior design, there is a growing need for technologies that analyze these design elements accurately and efficiently. To address this need, this study suggests a deep learning-based design analysis system. The proposed system consists of a semantic segmentation model that classifies spatial components and an image classification model that classifies attributes such as color, pattern, and material from the segmented components. Semantic segmentation model was trained using a dataset of 30000 personal indoor interior images collected for research, and during inference, the model separate the input image pixel into 34 categories. And experiments were conducted with various backbones in order to obtain the optimal performance of the deep learning model for the collected interior dataset. Finally, the model achieved good performance of 89.05% and 0.5768 in terms of accuracy and mean intersection over union (mIoU). In classification part convolutional neural network (CNN) model which has recorded high performance in other image recognition tasks was used. To improve the performance of the classification model we suggests an approach that how to handle data that has data imbalance and vulnerable to light intensity. Using our methods, we achieve satisfactory results in classifying interior design component attributes. In this paper, we propose indoor space design analysis system that automatically analyzes and classifies the attributes of indoor images using a deep learning-based model. This analysis system, used as a core module in the A.I interior recommendation service, can help users pursuing self-interior design to complete their designs more easily and efficiently.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.18 no.6
• /
• pp.1109-1116
• /
• 2023

Object tracking is used to track a goal in a video sequence by using coordinate information provided as annotation in the first frame of the video. In this paper, we propose a tracking algorithm that combines deep features and region inference modules to improve object tracking accuracy. In order to obtain sufficient object information, a convolution neural network was designed with a Siamese network structure. For object region inference, the region proposal network and overlapping confidence module were applied and used for tracking. The performance of the proposed tracking algorithm was evaluated using the Object Tracking Benchmark dataset, and it achieved 69.1% in the Success index and 89.3% in the Precision Metrics.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.26 no.6
• /
• pp.842-849
• /
• 2022

With the development of deep learning technology, there are many cases of using DNNs in embedded systems such as unmanned vehicles, drones, and robotics. Typically, in the case of an autonomous driving system, it is crucial to run several DNNs which have high accuracy results and large computation amount at the same time. However, running multiple DNNs simultaneously in an embedded system with relatively low performance increases the time required for the inference. This phenomenon may cause a problem of performing an abnormal function because the operation according to the inference result is not performed in time. To solve this problem, the solution proposed in this paper first reduces the computation by applying the Tucker decomposition to DNN models with big computation amount, and then, make DNN models run in parallel as much as possible in the unit of hidden layer inside the GPU. The experimental result shows that the DNN inference time decreases by up to 75.6% compared to the case before applying the proposed technique.

• Journal of Industrial Convergence
• /
• v.20 no.4
• /
• pp.23-30
• /
• 2022

Recently, with the development of artificial intelligence (AI) and deep learning, the importance of conversational artificial intelligence chatbots is being highlighted. In addition, chatbot research is being conducted in various fields. To build a chatbot, it is developed using an open source platform or a commercial platform for ease of development. These chatbot platforms mainly use RNN and application algorithms. The RNN algorithm has the advantages of fast learning speed, ease of monitoring and verification, and good inference performance. In this paper, a method for improving the inference performance of RNNs and applied algorithms was studied. The proposed method used the word group expansion learning technique of key words for each sentence when RNN and applied algorithm were applied. As a result of this study, the RNN, GRU, and LSTM three algorithms with a cyclic structure achieved a minimum of 0.37% and a maximum of 1.25% inference performance improvement. The research results obtained through this study can accelerate the adoption of artificial intelligence chatbots in related industries. In addition, it can contribute to utilizing various RNN application algorithms. In future research, it will be necessary to study the effect of various activation functions on the performance improvement of artificial neural network algorithms.

• Journal of Broadcast Engineering
• /
• v.27 no.3
• /
• pp.341-350
• /
• 2022

In this paper, we propose a deep learning based on-device augmented reality (AR) system in which multiple input images are used to implement the correct occlusion in a real environment. The proposed system is composed of three technical steps; camera pose estimation, depth estimation, and object augmentation. Each step employs various mobile frameworks to optimize the processing on the on-device environment. Firstly, in the camera pose estimation stage, the massive computation involved in feature extraction is parallelized using OpenCL which is the GPU parallelization framework. Next, in depth estimation, monocular and multiple image-based depth image inference is accelerated using the mobile deep learning framework, i.e. TensorFlow Lite. Finally, object augmentation and occlusion handling are performed on the OpenGL ES mobile graphics framework. The proposed augmented reality system is implemented as an application in the Android environment. We evaluate the performance of the proposed system in terms of augmentation accuracy and the processing time in the mobile as well as PC environments.

• Journal of Broadcast Engineering
• /
• v.25 no.2
• /
• pp.200-207
• /
• 2020

Recently, deep neural networks such as CNN are showing excellent performance in various fields such as image classification, object recognition, visual quality enhancement, etc. However, as the model size and computational complexity of deep learning models for most applications increases, it is hard to apply neural networks to IoT and mobile environments. Therefore, neural network compression algorithms for reducing the model size while keeping the performance have been being studied. In this paper, we apply few compression methods to CNN models and evaluate their performances in the embedded environment. For evaluate the performance, the classification performance and inference time of the original CNN models and the compressed CNN models on the image inputted by the camera are evaluated in the embedded board equipped with QCS605, which is a customized AI chip. In this paper, a few CNN models of MobileNetV2, ResNet50, and VGG-16 are compressed by applying the methods of pruning and matrix decomposition. The experimental results show that the compressed models give not only the model size reduction of 1.3~11.2 times at a classification performance loss of less than 2% compared to the original model, but also the inference time reduction of 1.2~2.21 times, and the memory reduction of 1.2~3.8 times in the embedded board.