• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2021.06a
• /
• pp.226-227
• /
• 2021

딥러닝을 활용한 영상 분석 기술은 GPU 하드웨어의 발전으로 인하여 소프트웨어 기반 처리 기술이 급격히 발전하였고 기존 패턴 분석 기술 대비 높은 정확도를 보여주고 있다. PC나 특정 하드웨어에서 동작하는 소프트웨어 기반 영상분석기술은 적용분야의 한계가 발생하였다. 신경망 기술을 하드웨어로 구현한 NPU(Network processing unit)의 개발로 고가의 플랫폼이 아닌 임베디드 플랫폼에서의 딥러닝 구현이 가능해졌다. 반면에 하드웨어에서 활용 가능한 네트워크가 제한적임으로 인하여 구현 가능한 딥러닝 모델의 크기, 메모리 등의 한계가 있으며 시시각각 변하는 딥러닝 기술에 기반한 최신모델 또는 고성능 모델을 구동하기에는 한계가 발생하였다. 이를 해결하기 위하여 본 연구에서는 Distillation 기법을 적용한 임베디드 시스템을 개발하고 이에 기반한 딥러닝 모델의 구현 및 상황에 따른 가변적 딥러닝 모델의 적용이 가능한 시스템을 구현하였다.

• Phonetics and Speech Sciences
• /
• v.13 no.3
• /
• pp.71-78
• /
• 2021

Most neural-network-based speech synthesis models utilize neural vocoders to convert mel-scaled spectrograms into high-quality, human-like voices. However, neural vocoders combined with mel-scaled spectrogram prediction models demand considerable computer memory and time during the training phase and are subject to slow inference speeds in an environment where GPU is not used. This problem does not arise in linear spectrogram prediction models, as they do not use neural vocoders, but these models suffer from low voice quality. As a solution, this paper proposes a Tacotron 2 and Transformer-based linear spectrogram prediction model that produces high-quality speech and does not use neural vocoders. Experiments suggest that this model can serve as the foundation of a high-quality text-to-speech model with fast inference speed.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.26 no.6
• /
• pp.850-858
• /
• 2022

To deploy Gate Recurrent Units (GRU) on resource-constrained embedded devices, this paper presents a reconfigurable FPGA-based GRU accelerator that enables structured compression. Firstly, a dense GRU model is significantly reduced in size by hybrid quantization and structured top-k pruning. Secondly, the energy consumption on external memory access is greatly reduced by the proposed reuse computing pattern. Finally, the accelerator can handle a structured sparse model that benefits from the algorithm-hardware co-design workflows. Moreover, inference tasks can be flexibly performed using all functional dimensions, sequence length, and number of layers. Implemented on the Intel DE1-SoC FPGA, the proposed accelerator achieves 45.01 GOPs in a structured sparse GRU network without batching. Compared to the implementation of CPU and GPU, low-cost FPGA accelerator achieves 57 and 30x improvements in latency, 300 and 23.44x improvements in energy efficiency, respectively. Thus, the proposed accelerator is utilized as an early study of real-time embedded applications, demonstrating the potential for further development in the future.

• Annual Conference on Human and Language Technology
• /
• 2021.10a
• /
• pp.259-263
• /
• 2021

상호참조해결은 주어진 문서에서 상호참조해결 대상이 되는 멘션(mention)을 식별하고, 동일한 개체(entity)를 의미하는 멘션들을 찾아 그룹화하는 자연어처리 태스크이다. 최근 상호참조해결에서는 BERT를 이용하여 단어의 문맥 표현을 얻은 후, 멘션 탐지와 상호참조해결을 동시에 진행하는 end-to-end 모델이 주로 연구되었으나, 512 토큰 이상의 긴 문서를 처리하기 위해서는 512 토큰 이하로 문서를 분할하여 처리하기 때문에 길이가 긴 문서에 대해서는 상호참조해결 성능이 낮아지는 문제가 있다. 본 논문에서는 512 토큰 이상의 긴 문서를 위한 BERT 기반의 end-to-end 상호참조해결 모델을 제안한다. 본 모델은 긴 문서를 512 이하의 토큰으로 쪼개어 기존의 BERT에서 단어의 1차 문맥 표현을 얻은 후, 이들을 다시 연결하여 긴 문서의 Global Positional Encoding 또는 Embedding 값을 더한 후 Global BERT layer를 거쳐 단어의 최종 문맥 표현을 얻은 후, end-to-end 상호참조해결 모델을 적용한다. 실험 결과, 본 논문에서 제안한 모델이 기존 모델과 유사한 성능을 보이면서(테스트 셋에서 0.16% 성능 향상), GPU 메모리 사용량은 1.4배 감소하고 속도는 2.1배 향상되었다.

• KIPS Transactions on Computer and Communication Systems
• /
• v.5 no.8
• /
• pp.181-188
• /
• 2016

Recently, as the amount of spatial information increases, an interest in the study of spatial information processing has been increased. Spatial database systems extended from the traditional relational database systems are difficult to handle large data sets because of the scalability. SpatialHadoop extended from Hadoop system has a low performance, because spatial computations in SpationHadoop require a lot of write operations of intermediate results to the disk, resulting in the performance degradation. In this paper, Spatial Computation Spark(SC-Spark) is proposed, which is an in-memory based distributed processing framework. SC-Spark is extended from Spark in order to efficiently perform the spatial operation for large-scale data. In addition, SC-Spark based on the GPGPU is developed to improve the performance of the SC-Spark. SC-Spark uses the advantage of the Spark holding intermediate results in the memory. And GPGPU-based SC-Spark can perform spatial operations in parallel using a plurality of processing elements of an GPU. To verify the proposed work, experiments on a single AMD system were performed using SC-Spark and GPGPU-based SC-Spark for Point-in-Polygon and spatial join operation. The experimental results showed that the performance of SC-Spark and GPGPU-based SC-Spark were up-to 8 times faster than SpatialHadoop.

• Journal of Digital Contents Society
• /
• v.16 no.5
• /
• pp.805-813
• /
• 2015

Many researchers expect the compressive sensing and sparse recovery problem can overcome the limitation of conventional digital techniques. However, these new approaches require to solve the l1 norm optimization problems when it comes to signal reconstruction. In the signal reconstruction process, the transform computation by multiplication of a random matrix and a vector consumes considerable computing power. To address this issue, parallel processing is applied to the optimization problems. In particular, due to huge size of original signal, it is hard to store the random matrix directly in memory, which makes one need to design a procedural approach in handling the random matrix. This paper presents a new parallel algorithm to calculate random partial Haar wavelet transform based on Single Instruction Multiple Threads (SIMT) platform.

• Journal of Intelligence and Information Systems
• /
• v.28 no.1
• /
• pp.89-106
• /
• 2022

Recently, the amount of video data collected from smartphones, CCTVs, black boxes, and high-definition cameras has increased rapidly. According to the increasing video data, the requirements for analysis and utilization are increasing. Due to the lack of skilled manpower to analyze videos in many industries, machine learning and artificial intelligence are actively used to assist manpower. In this situation, the demand for various computer vision technologies such as object detection and tracking, action detection, emotion detection, and Re-ID also increased rapidly. However, the object detection and tracking technology has many difficulties that degrade performance, such as re-appearance after the object's departure from the video recording location, and occlusion. Accordingly, action and emotion detection models based on object detection and tracking models also have difficulties in extracting data for each object. In addition, deep learning architectures consist of various models suffer from performance degradation due to bottlenects and lack of optimization. In this study, we propose an video analysis system consists of YOLOv5 based DeepSORT object tracking model, SlowFast based action recognition model, Torchreid based Re-ID model, and AWS Rekognition which is emotion recognition service. Proposed model uses single-linkage hierarchical clustering based Re-ID and some processing method which maximize hardware throughput. It has higher accuracy than the performance of the re-identification model using simple metrics, near real-time processing performance, and prevents tracking failure due to object departure and re-emergence, occlusion, etc. By continuously linking the action and facial emotion detection results of each object to the same object, it is possible to efficiently analyze videos. The re-identification model extracts a feature vector from the bounding box of object image detected by the object tracking model for each frame, and applies the single-linkage hierarchical clustering from the past frame using the extracted feature vectors to identify the same object that failed to track. Through the above process, it is possible to re-track the same object that has failed to tracking in the case of re-appearance or occlusion after leaving the video location. As a result, action and facial emotion detection results of the newly recognized object due to the tracking fails can be linked to those of the object that appeared in the past. On the other hand, as a way to improve processing performance, we introduce Bounding Box Queue by Object and Feature Queue method that can reduce RAM memory requirements while maximizing GPU memory throughput. Also we introduce the IoF(Intersection over Face) algorithm that allows facial emotion recognized through AWS Rekognition to be linked with object tracking information. The academic significance of this study is that the two-stage re-identification model can have real-time performance even in a high-cost environment that performs action and facial emotion detection according to processing techniques without reducing the accuracy by using simple metrics to achieve real-time performance. The practical implication of this study is that in various industrial fields that require action and facial emotion detection but have many difficulties due to the fails in object tracking can analyze videos effectively through proposed model. Proposed model which has high accuracy of retrace and processing performance can be used in various fields such as intelligent monitoring, observation services and behavioral or psychological analysis services where the integration of tracking information and extracted metadata creates greate industrial and business value. In the future, in order to measure the object tracking performance more precisely, there is a need to conduct an experiment using the MOT Challenge dataset, which is data used by many international conferences. We will investigate the problem that the IoF algorithm cannot solve to develop an additional complementary algorithm. In addition, we plan to conduct additional research to apply this model to various fields' dataset related to intelligent video analysis.