• KIPS Transactions on Computer and Communication Systems
• /
• v.9 no.12
• /
• pp.291-306
• /
• 2020

Nowadays, Data-Network-AI (DNA)-based intelligent services and applications have become a reality to provide a new dimension of services that improve the quality of life and productivity of businesses. Artificial intelligence (AI) can enhance the value of IoT data (data collected by IoT devices). The internet of things (IoT) promotes the learning and intelligence capability of AI. To extract insights from massive volume IoT data in real-time using deep learning, processing capability needs to happen in the IoT end devices where data is generated. However, deep learning requires a significant number of computational resources that may not be available at the IoT end devices. Such problems have been addressed by transporting bulks of data from the IoT end devices to the cloud datacenters for processing. But transferring IoT big data to the cloud incurs prohibitively high transmission delay and privacy issues which are a major concern. Edge computing, where distributed computing nodes are placed close to the IoT end devices, is a viable solution to meet the high computation and low-latency requirements and to preserve the privacy of users. This paper provides a comprehensive review of the current state of leveraging deep learning within edge computing to unleash the potential of IoT big data generated from IoT end devices. We believe that the revision will have a contribution to the development of DNA-based intelligent services and applications. It describes the different distributed training and inference architectures of deep learning models across multiple nodes of the edge computing platform. It also provides the different privacy-preserving approaches of deep learning on the edge computing environment and the various application domains where deep learning on the network edge can be useful. Finally, it discusses open issues and challenges leveraging deep learning within edge computing.

• International Journal of Advanced Culture Technology
• /
• v.8 no.1
• /
• pp.165-172
• /
• 2020

Industrial Revolution means the improvement of productivity through technological innovation and has been a driving force of the whole change of economic system and social structure as the characteristic of technology as the tool of this productivity has changed. Since the first industrial revolution of the 18th century, productivity efficiency has been advanced through three industrial revolutions so far, and this fourth industrial revolution is expected to bring about another revolution of production. In this study, the demand for the introduction of artificial intelligence(AI) technology has been increasing in various business fields due to the rapid development of ICT technology, and the classification of HS(harmonized commodity description and coding system) items has been decided using artificial intelligence technology, which is the core of the fourth industrial revolution. And it is enough to construct HS classification system based on AI technology using inference and deep learning. Performing the HS item classification is not an easy task. Implementation of item classification system using artificial intelligence technology to analyze information of HS item classification which is performed manually by the current person more accurately and without any mistake, And the customs administrations, customs offices, and customs agencies, it is expected to be highly utilized in the innovation of trade practice and the customs administration innovation FTA origin agent.

• ETRI Journal
• /
• v.40 no.4
• /
• pp.421-434
• /
• 2018

The application of deep neural networks (DNNs) to connect the world with cyber physical systems (CPSs) has attracted much attention. However, DNNs require a large amount of memory and computational cost, which hinders their use in the relatively low-end smart devices that are widely used in CPSs. In this paper, we aim to determine whether DNNs can be efficiently deployed and operated in low-end smart devices. To do this, we develop a method to reduce the memory requirement of DNNs and increase the inference speed, while maintaining the performance (for example, accuracy) close to the original level. The parameters of DNNs are decomposed using a hybrid of canonical polyadic-singular value decomposition, approximated using a tensor power method, and fine-tuned by performing iterative one-shot hybrid fine-tuning to recover from a decreased accuracy. In this study, we evaluate our method on frequently used networks. We also present results from extensive experiments on the effects of several fine-tuning methods, the importance of iterative fine-tuning, and decomposition techniques. We demonstrate the effectiveness of the proposed method by deploying compressed networks in smartphones.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2021.05a
• /
• pp.538-540
• /
• 2021

In this paper, we proposed a smart target detection system that detects and recognizes a designated target to provide relative motion information when performing a target detection mission of a drone. The proposed system focused on developing an algorithm that can secure adequate accuracy (i.e. mAP, IoU) and high real-time at the same time. The proposed system showed an accuracy of close to 1.0 after 100k learning of the Google Inception V2 deep learning model, and the inference speed was about 60-80[Hz] when using a high-performance laptop based on the real-time performance Nvidia GTX 2070 Max-Q. The proposed smart target detection system will be operated like a drone and will be helpful in successfully performing surveillance and reconnaissance missions by automatically recognizing the target using computer image processing and following the target.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.05a
• /
• pp.123-125
• /
• 2023

As deep learning has become an essential part of human lives, the requirement for Deep Learning as a Service (DLaaS) is growing. Since using remote cloud servers induces privacy concerns for users, a Fully Homomorphic Encryption (FHE) arises to protect users' sensitive data from a malicious attack in the cloud environment. However, the FHE cannot support several computations, including the most popular activation function, Rectified Linear Unit (ReLU). This paper analyzes several polynomial approximation methods for ReLU to utilize FHE in DLaaS.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.33 no.6
• /
• pp.861-868
• /
• 2023

One of the key technologies in providing data analysis in the deep learning while maintaining security is fully homomorphic encryption. Due to constraints in operations on fully homomorphically encrypted data, non-arithmetic functions used in deep learning must be approximated by polynomials. Until now, the degrees of approximation polynomials with composite minimax polynomials have been uniformly set across layers, which poses challenges for effective network designs on fully homomorphic encryption. This study theoretically proves that setting different degrees of approximation polynomials constructed by composite minimax polynomial in each layer does not pose any issues in the inference on convolutional neural networks.

• Journal of Bio-Environment Control
• /
• v.31 no.3
• /
• pp.255-260
• /
• 2022

This study aimed to develop a service model that uses a deep learning algorithm for detecting diseases and pests in strawberries through image data. In addition, the pest detection performance of deep learning models was further improved by proposing segmented image data sets specialized in disease and pest symptoms. The CNN-based YOLO deep learning model was selected to enhance the existing R-CNN-based model's slow learning speed and inference speed. A general image data set and a proposed segmented image dataset was prepared to train the pest and disease detection model. When the deep learning model was trained with the general training data set, the pest detection rate was 81.35%, and the pest detection reliability was 73.35%. On the other hand, when the deep learning model was trained with the segmented image dataset, the pest detection rate increased to 91.93%, and detection reliability was increased to 83.41%. This study concludes with the possibility of improving the performance of the deep learning model by using a segmented image dataset instead of a general image dataset.

• Journal of Broadcast Engineering
• /
• v.23 no.4
• /
• pp.503-510
• /
• 2018

Restoring of damaged images is a fundamental problem that was attempted before digital image processing technology appeared. Various algorithms for reconstructing damaged images have been introduced. However, the results show inferior restoration results compared with manual restoration. Recent developments of DNN (Deep Neural Network) have introduced various studies that apply it to image restoration. However, if the wide area is damaged, it can not be solved by a general interpolation method. In this case, it is necessary to reconstruct the damaged area through contextual information of surrounding images. In this paper, we propose an image restoration network using a generative adversarial network (GAN). The proposed system consists of image generation network and discriminator network. The proposed network is verified through experiments that it is possible to recover not only the natural image but also the texture of the original image through the inference of the damaged area in restoring various types of images.

• Journal of Broadcast Engineering
• /
• v.25 no.2
• /
• pp.275-278
• /
• 2020

In this paper, we proposed a deep learning based character recognition system through virtual data generation. In order to secure the learning data that takes the largest weight in supervised learning, virtual data was created. Also, after creating virtual data, data generalization was performed to cope with various data by using augmentation parameter. Finally, the learning data composition generated data by assigning various values to augmentation parameter and font parameter. Test data for measuring the character recognition performance was constructed by cropping the text area from the actual image data. The test data was augmented considering the image distortion that may occur in real environment. Deep learning algorithm uses YOLO v3 which performs detection in real time. Inference result outputs the final detection result through post-processing.

• Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography
• /
• v.39 no.6
• /
• pp.381-392
• /
• 2021

In this study, we perform deep learning-based object detection analysis on eight types of buildings defined by the digital map topography standard code, leveraging images taken with UAV (Unmanned Aerial Vehicle). Image labeling was done for 509 images taken by UAVs and the YOLO (You Only Look Once) v5 model was applied to proceed with learning and inference. For experiments and analysis, data were analyzed by applying an open source-based analysis platform and algorithm, and as a result of the analysis, building objects were detected with a prediction probability of 88% to 98%. In addition, the learning method and model construction method necessary for the high accuracy of building object detection in the process of constructing and repetitive learning of training data were analyzed, and a method of applying the learned model to other images was sought. Through this study, a model in which high-efficiency deep neural networks and spatial information data are fused will be proposed, and the fusion of spatial information data and deep learning technology will provide a lot of help in improving the efficiency, analysis and prediction of spatial information data construction in the future.