• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2014.06a
• /
• pp.157-159
• /
• 2014

In this paper, we introduce SlimSRP, an ultra-low-power digital signal processor (DSP) solution for mobile audio and voice applications. So far, application processors (APs) have taken charge of all the tasks in mobile devices. However, they have suffered from short battery life problems to deal with complex usage scenarios, such as always-on voice trigger with continuous audio playback. From extensive analysis of audio and voice application characteristics, SlimSRP is designed to relive the performance and power burden of APs. It employs three-issue VLIW architecture, and the major low-power and high-performance techniques include: (1) an optimized register-file architecture friendly for constants generation, (2) a powerful instruction set to reduce the number of register file accesses and (3) a unique instruction compression scheme that contributes to saved memory size and reduced cache miss. An implementation of SlimSRP runs at up to 200MHz and the logic occupies 95K NAND2 gates in Samsung 28LPP process. The experimental results demonstrate that a MP3 decoder application with a 128kbps 44.1kHz input can run at 5.1MHz and the logic consumes only 22uW/MHz.

• Journal of Broadcast Engineering
• /
• v.16 no.4
• /
• pp.596-601
• /
• 2011

In this paper, we propose a fast depth-image-based rendering method to generate a virtual view image in real-time using a graphic processor unit (GPU) for a 3D broadcasting system. Before the transmission, we encode the input 2D+depth video using the H.264 coding standard. At the receiver, we decode the received bitstream and generate a stereo video using a GPU which can compute in parallel. In this paper, we apply a simple and efficient hole filling method to reduce the decoder complexity and reduce hole filling errors. Besides, we design a vertical parallel structure for a forward mapping process to take advantage of the single instruction multiple thread structure of GPU. We also utilize high speed GPU memories to boost the computation speed. As a result, we can generate virtual view images 15 times faster than the case of CPU-based processing.

• The KIPS Transactions:PartA
• /
• v.18A no.3
• /
• pp.109-114
• /
• 2011

We describe a technique for automatically proving compiler optimizations sound, meaning that their transformations are always semantics-preserving. As is well known, IR (Intermediate Representation) optimization is an important step in a compiler backend. But unfortunately, it is difficult to detect and debug the IR optimization errors for compiler developers. So, we introduce a C level error check system for detecting the correctness of these IR transformation techniques. In our system, we first create an IR-to-C converter to translate IR to C code before and after each compiler optimization phase, respectively, since our technique is based on the Memory Comparison-based Clone(MeCC) detector which is a tool of detecting semantic equivalency in C level. MeCC accepts only C codes as its input and it uses a path-sensitive semantic-based static analyzer to estimate the memory states at exit point of each procedure, and compares memory states to determine whether the procedures are equal or not. But MeCC cannot guarantee two semantic-equivalency codes always have 100% similarity or two codes with different semantics does not get the result of 100% similarity. To increase the reliability of the results, we describe a technique which comprises how to generate C codes in IR-to-C transformation phase and how to send the optimization information to MeCC to avoid the occurrence of these unexpected problems. Our methodology is illustrated by three familiar optimizations, dead code elimination, instruction scheduling and common sub-expression elimination and our experimental results show that the C level error check system is highly reliable.

• KIPS Transactions on Software and Data Engineering
• /
• v.8 no.9
• /
• pp.379-390
• /
• 2019

In this paper, we propose a novel deep neural network model for Vision-and-Language Navigation (VLN) named LVLN (Landmark-based VLN). In addition to both visual features extracted from input images and linguistic features extracted from the natural language instructions, this model makes use of information about places and landmark objects detected from images. The model also applies a context-based attention mechanism in order to associate each entity mentioned in the instruction, the corresponding region of interest (ROI) in the image, and the corresponding place and landmark object detected from the image with each other. Moreover, in order to improve the success rate of arriving the target goal, the model adopts a progress monitor module for checking substantial approach to the target goal. Conducting experiments with the Matterport3D simulator and the Room-to-Room (R2R) benchmark dataset, we demonstrate high performance of the proposed model.

• Asia-pacific Journal of Multimedia Services Convergent with Art, Humanities, and Sociology
• /
• v.7 no.3
• /
• pp.339-349
• /
• 2017

This paper proposes a four-stage algorithm for detecting lanes on a driving car. In the first stage, it extracts region of interests in an image. In the second stage, it employs a median filter to remove noise. In the third stage, a binary algorithm is used to classify two classes of backgrond and foreground of an input image. Finally, an image erosion algorithm is utilized to obtain clear lanes by removing noises and edges remained after the binary process. However, the proposed lane detection algorithm requires high computational time. To address this issue, this paper presents a parallel implementation of a real-time line detection algorithm on a multi-core architecture. In addition, we implement and simulate 8 different processing element (PE) architectures to select an optimal PE architecture for the target application. Experimental results indicate that 40×40 PE architecture show the best performance, energy efficiency and area efficiency.

• KIPS Transactions on Software and Data Engineering
• /
• v.12 no.11
• /
• pp.481-492
• /
• 2023

This study explores how to build a Korean dataset to extract information from text using generative large language models. In modern society, mixed information circulates rapidly, and effectively categorizing and extracting it is crucial to the decision-making process. However, there is still a lack of Korean datasets for training. To overcome this, this study attempts to extract information using text-based zero-shot learning using a generative large language model to build a purposeful Korean dataset. In this study, the language model is instructed to output the desired result through prompt engineering in the form of "system"-"instruction"-"source input"-"output format", and the dataset is built by utilizing the in-context learning characteristics of the language model through input sentences. We validate our approach by comparing the generated dataset with the existing benchmark dataset, and achieve 25.47% higher performance compared to the KLUE-RoBERTa-large model for the relation information extraction task. The results of this study are expected to contribute to AI research by showing the feasibility of extracting knowledge elements from Korean text. Furthermore, this methodology can be utilized for various fields and purposes, and has potential for building various Korean datasets.