ETRI Journal最新文献

When reconstructing extensive terrain, it is essential to partition it into smaller tiles for individual processing. This paper introduces a texture reconstruction approach that ensures seamless and consistent final outputs, even when processed tile by tile. Among the stages of multi-view image-based reconstruction, texture reconstruction presents significant challenges during tile-based processing. Relying solely on local tile-level data complicates achieving precise texture mapping. The absence of occlusion details between tiles can lead to selecting incorrect images as the best visible ones or adjusting tile texture colors differently, resulting in noticeable grid-like texture seams in the final result. To mitigate these issues, we leverage global depth maps to accurately detect occlusions between neighboring tiles. Furthermore, by utilizing a shared texture candidate list, we establish uniform targets for texture color correction across tiles. Experimental findings demonstrate that leveraging global information for texture reconstruction on a tile-by-tile basis enables the creation of smooth and realistic texture maps, as validated through comparisons with existing methodologies.

Network function parallelism (NFP) has gained attention for processing packets in parallel through service functions arranged in the required service function chain. While parallel processing efficiently reduces the service function chaining (SFC) completion time, it incurs a higher network overhead (e.g., network congestion) to replicate various packets for processing. To reduce the SFC completion time while maintaining a low network overhead, we propose a deep-reinforcement-learning-based NFP algorithm (DeepNFP) that provides an SFC processing policy to determine the sequential or parallel processing of every service function. In DeepNFP, deep reinforcement learning captures the network dynamics and service function conditions and iteratively finds the SFC processing policy in the network environment. Furthermore, an SFC data plane protocol based on segment routing over IPv6 configures and operates the policy in the SFC data network. Evaluation results show that DeepNFP can achieve 46% of the SFC completion time and 66% of the network overhead compared with conventional SFC and NFP, respectively.

Semantic segmentation is essential in machine vision but susceptible to noise and distortions that often appear in real-world images. We propose UPlus-Net (UP-Net), a deep-learning architecture based on the U-Net encoder–decoder architecture. We address the limitations of U-Net by introducing a multi-head architecture in UP-Net to properly handle segmentation challenges. In addition, we evaluate UP-Net for decoding distorted quick-response (QR) codes heavily polluted by noise. Experimental results confirm that UP-Net outperforms existing QR reader mobile applications, highlighting the UP-Net ability to handle challenging images. Unlike existing methods focused solely on QR code reading or segmentation, UP-Net offers a combined solution, efficiently and accurately reading distorted QR codes while performing high-quality semantic segmentation. These unique characteristics render UP-Net promising for applications demanding robust image analysis in challenging environments.

In the digital security environment, the obfuscation and encryption of malicious scripts are primary attack methods used to evade detection. These scripts—easily spread through websites, emails, and file downloads—can be automatically executed on users' systems, posing serious security threats. To overcome the limitations of signature-based detection methods, this study proposed a methodology for real-time detection of obfuscated and encrypted malicious scripts using ML/DL models with feature optimization techniques. The obfuscated script datasets were analyzed to identify the unique characteristics, classified into 16 feature sets, to evaluate the optimal features for the best detection accuracy. Although the detection accuracy of these datasets was < 20%, when tested with commercial antivirus services, the experimental results using ML and DL models demonstrated that the proposed light gradient boosting model (LGBM) could achieve the best detection accuracy and processing speed. The LGBM outperformed other artificial intelligence models by achieving 97% accuracy and the minimum processing time in the decoded, obfuscated, and encrypted dataset cases.

NEST-C: A deep learning compiler framework for heterogeneous computing systems with artificial intelligence accelerators

https://doi.org/10.4218/etrij.2024-0139

ETRI Journal, Volume 46, Issue 5, October 2024, pp. 851–864.

In the article entitled “NEST-C: A deep learning compiler framework for heterogeneous computing systems with artificial intelligence accelerators,” the authors would like to correct the funding information of their article. It should be written as follows:

Funding information This study is supported by a grant from the Institute of Information & Communications Technology Planning & Evaluation (IITP), funded by the Korean government (MSIT) (No. RS-2023-00277060, Development of OpenEdge AI SoC hardware and software platform and No. 2018-0-00769, Neuromorphic Computing Software Platform for Artificial Intelligence Systems).

The authors would like to apologize for the inconvenience caused.

Sungryeul Rhyu | Junsik Kim | Gwang Hoon Park | Kyuheon Kim

Low-complexity patch projection method for efficient and lightweight point-cloud compression

https://doi.org/10.4218/etrij.2023-0242

ETRI Journal, Volume 46, Issue 4, August 2024, pp. 683–696.

In the article entitled “Low-complexity patch projection method for efficient and lightweight point-cloud compression”, the authors would like to correct the funding information of their article. It should be written as follows:

Funding information

This study was supported by the Information Technology Research Center of the Ministry of Science and ICT, Korea (grant number: IITP-2024-2021-0-02046) and the Institute of Information & Communications Technology Planning & Evaluation, Korea (grant number: RS-2023-00227431, Development of 3D space digital media standard technology).

The authors would like to apologize for the inconvenience caused.

Deep-learning methods, such as encoder–decoder networks, have achieved impressive results in image dehazing. However, these methods often rely only on synthesized data for training that limits their generalizability to hazy, real-world images. To leverage prior knowledge of haze properties, we propose a bi-encoder structure that integrates a prior-based encoder into a traditional encoder–decoder network. The features from both encoders were fused using a feature enhancement module. We adopted transformer blocks instead of convolutions to model local feature associations. Experimental results demonstrate that our method surpasses state-of-the-art methods for synthesized and actual hazy scenes. Therefore, we believe that our method will be a useful supplement to the collection of current artificial intelligence models and will benefit engineering applications in computer vision.

The application of the Internet of Things generates massive end-device data. Traditional end-to-edge data caching methods lack parallelism, consume numerous resources, and lack data backup mechanisms. We propose a lightweight high-speed data caching and forwarding architecture based on a field-programmable gate array. In the basic scheme, a double-data-rate memory read/write mixed control method is proposed to efficiently cache massive amounts of data. To improve reliability, we propose address-space virtualization to forward data to backup devices. To satisfy the demand for higher performance scenarios, we propose a real-time data shunting mechanism to realize an optimized scheme that supports higher data rates. These two schemes support single-link high-speed data caching and forwarding at 10 and 40 Gbps. Compared with the existing method, the basic and optimized schemes improve the performance by 25% and four times and save 66% and 40% of the resources, respectively. Thus, we provide a lightweight and reliable high-performance end-to-edge data caching solution for resource-constrained edge devices.

SoCs with analog-circuit-based unsigned weight-accumulating spiking neural networks (UWA-SNNs) are a highly promising solution for achieving low-power AI-SoCs. This paper addresses the challenges that must be overcome to realize the potential of UWA-SNNs in low-power AI-SoCs: (i) the absence of UWA-SNN learning methods and the lack of an environment for developing applications based on trained SNN models and (ii) the inherent issue of testing and validating applications on the system being nearly impractical until the final chip is fabricated owing to the mixed-signal circuit implementation of UWA-SNN-based SoCs. This paper argues that, by integrating the proposed solutions, the development of an EDA tool that enables the easy and rapid development of UWA-SNN-based SoCs is feasible, and demonstrates this through the development of the SNN eXpress (SNX) tool. The developed SNX automates the generation of RTL code, FPGA prototypes, and a software development kit tailored for UWA-SNN-based application development. Comprehensive details of SNX development and the performance evaluation and verification results of two AI-SoCs developed using SNX are also presented.

Owing to the widespread advancement of transformer-based artificial neural networks, artificial intelligence (AI) processors are now required to perform matrix–vector multiplication in addition to the conventional matrix–matrix multiplication. However, current AI processor architectures are optimized for general matrix–matrix multiplications (GEMMs), which causes significant throughput degradation when processing general matrix–vector multiplications (GEMVs). In this study, we proposed a port-folding GEMV (PF-GEMV) scheme employing multiformat and low-precision techniques while reusing an outer product-based processor optimized for conventional GEMM operations. This approach achieves 93.7% utilization in GEMV operations with an 8-bit format on an 8 $�\times$ 8 processor, thus resulting in a 7.5 $�\times$ increase in throughput compared with that of the original scheme. Furthermore, when applied to the matrix operation of the GPT-2 large model, an increase in speed by 7 $�\times$ is achieved in single-batch inferences.