• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.11a
• /
• pp.695-697
• /
• 2023

최근 몇 년 간 딥러닝 기반 모델의 규모와 복잡성이 증가하면서 강력하고, 높은 정확도가 확보되지만 많은 양의 계산 자원과 메모리가 필요하기 때문에 모바일 장치나 임베디드 시스템과 같은 리소스가 제한된 환경에서의 배포에 제약사항이 생긴다. 복잡한 딥러닝 모델의 배포 및 운영 시 요구되는 고성능 컴퓨터 자원의 문제점을 해결하고자 사전 학습된 대규모 모델로부터 가벼운 모델을 학습시키는 지식증류 기법이 제안되었다. 하지만 현대 딥러닝 기반 모델은 높은 정확도 대비 훈련 데이터에 과적합 되는 과잉 확신(overconfidence) 문제에 대한 대책이 필요하다. 본 논문은 효율적인 경량화를 위한 미리 학습된 모델의 과잉 확신을 방지하고자 초점 손실(focal loss)을 이용한 모델 보정 기법을 언급하며, 다양한 손실 함수 변형에 따라서 지식증류의 성능이 어떻게 변화하는지에 대해 탐구하고자 한다.

• The Korean Journal of Applied Statistics
• /
• v.34 no.6
• /
• pp.979-993
• /
• 2021

Acoustic scene classification (ASC) categorizes an audio file based on the environment in which it has been recorded. This has long been studied in the detection and classification of acoustic scenes and events (DCASE). In this study, we considered the problem that ASC faces in real-world applications that the model used should have low-complexity. We compared several models that apply light-weight techniques. First, a base CNN model was proposed using log mel-spectrogram, deltas, and delta-deltas features. Second, depthwise separable convolution, linear bottleneck inverted residual block was applied to the convolutional layer, and Quantization was applied to the models to develop a low-complexity model. The model considering low-complexity was similar or slightly inferior to the performance of the base model, but the model size was significantly reduced from 503 KB to 42.76 KB.

• Journal of IKEEE
• /
• v.26 no.4
• /
• pp.639-646
• /
• 2022

In this paper, to solve the large dataset problem, we collect images through an image collection method called web crawling and build datasets for use in image classification models through a data preprocessing process. We also propose a lightweight model that can automatically classify images by adding category values by incorporating transfer learning into the image classification model and an image classification model that reduces training time and achieves high accuracy.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.21 no.1
• /
• pp.93-98
• /
• 2021

Due to the development of deep learning and AI, the scale of the model has grown, and it has been integrated into other fields to blend into our lives. However, in environments with limited resources such as embedded devices, it is exist difficult to apply the model and problems such as power shortages. To solve this, lightweight methods such as clouding or offloading technologies, reducing the number of parameters in the model, or optimising calculations are proposed. In this paper, quantization of learned models is applied to ONNX models used in various framework interchange formats, neural network structure and inference performance are compared with existing models, and various module methods for quantization are analyzed. Experiments show that the size of weight parameter is compressed and the inference time is more optimized than before compared to the original model.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• fall
• /
• pp.80-81
• /
• 2021

딥 러닝 모델 사용에 있어서, 일반적인 사용자가 이용할 수 있는 하드웨어 리소스는 제한적이기 때문에 기존 모델을 경량화 할 수 있는 프루닝 방법을 통해 제한적인 리소스를 효과적으로 활용할 수 있도록 한다. 그 방법으로, 여러 딥 러닝 모델들 중 비교적 파라미터 수가 많은 것으로 알려진 GAN 아키텍처에 네트워크 프루닝을 적용함으로써 비교적 무거운 모델을 적은 파라미터를 통해 학습할 수 있는 방법을 제시한다. 또한, 본 논문을 통해 기존의 SRGAN 논문에서 가장 효과적인 결과로 제시했던 16 개의 residual block 의 개수를 실제로 줄여 봄으로써 기존 논문에서 제시했던 결과와의 차이에 대해 서술한다.

• Journal of Convergence for Information Technology
• /
• v.11 no.9
• /
• pp.28-34
• /
• 2021

In the case of object detection using deep learning, both accuracy and real-time are required. However, it is difficult to use a deep learning model that processes a large amount of data in a limited resource environment. To solve this problem, this paper proposes an object detection model for small embedded devices. Unlike the general detection model, the model size was minimized by using a structure in which the pre-trained feature extractor was removed. The structure of the model was designed by repeatedly stacking lightweight convolution blocks. In addition, the number of region proposals is greatly reduced to reduce detection overhead. The proposed model was trained and evaluated using the public dataset PASCAL VOC. For quantitative evaluation of the model, detection performance was measured with average precision used in the detection field. And the detection speed was measured in a Raspberry Pi similar to an actual embedded device. Through the experiment, we achieved improved accuracy and faster reasoning speed compared to the existing detection method.

• Journal of Korea Society of Industrial Information Systems
• /
• v.28 no.4
• /
• pp.1-9
• /
• 2023

In this paper, we propose a method to recognize the braille blocks for embedded devices in real time through deep learning. First, a deep learning model for braille block recognition is trained on a high-performance computer, and the learning model is applied to a lightweight tool to apply to an embedded device. To recognize the walking information of the braille block, an algorithm is used to determine the path using the distance from the braille block in the image. After detecting braille blocks, bollards, and crosswalks through the YOLOv8 model in the video captured by the embedded device, the walking information is recognized through the braille block path discrimination algorithm. We apply the model lightweight tool to YOLOv8 to detect braille blocks in real time. The precision of YOLOv8 model weights is lowered from the existing 32 bits to 8 bits, and the model is optimized by applying the TensorRT optimization engine. As the result of comparing the lightweight model through the proposed method with the existing model, the path recognition accuracy is 99.05%, which is almost the same as the existing model, but the recognition speed is reduced by 59% compared to the existing model, processing about 15 frames per second.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2024.05a
• /
• pp.719-720
• /
• 2024

센서 기반의 자동화재탐지설비의 역할을 지원할 목적으로, 합성곱 신경망 기반의 AI 화재 감시장비등이 연구되어왔다. ai 기반 화재 감지에 사용되는 알고리즘은 전이학습을 주로 이용하고 있고, 이는 화재 감지에 기여도가 낮은 프로세스가 내장되어 있을 가능성이 존재하여, 딥러닝 모델의 복잡성을 가중시키는 원인이 될 수 있다. 본 연구에서는 이러한 모델의 복잡성을 개선하고자 다양한 딥러닝 및 해석 기술들을 분석하였고, 분석 결과를 토대로 화재 감지에 최적화된 아키텍처인 "IF2bNet" 을 제안한다. 구현한 아키텍처의 성능을 비교한 결과 동일한 성능을 내면서, 파라미터를 약 0.1 배로 경량화 하여, 복잡성을 완화하였다.

• Journal of IKEEE
• /
• v.28 no.2
• /
• pp.168-173
• /
• 2024

Speaker recognition refers to a technology that analyzes voice frequencies that are different for each individual and compares them with pre-stored voices to determine the identity of the person. Deep learning-based speaker recognition is being applied to many fields, and pet robots are one of them. However, the hardware performance of pet robots is very limited in terms of the large memory space and calculations of deep learning technology. This is an important problem that pet robots must solve in real-time interaction with users. Lightening deep learning models has become an important way to solve the above problems, and a lot of research is being done recently. In this paper, we describe the results of research on lightweight speaker recognition for pet robots by constructing a voice data set for pet robots, which is a specific command type, and comparing the results of models using residuals. In the conclusion, we present the results of the proposed method and Future research plans are described.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• fall
• /
• pp.152-153
• /
• 2021

고성능 딥러닝 모델은 학습과 추론 과정에서 고비용의 전산 자원과 많은 연산량을 필요로 하여 이에 따른 개발 환경과 많은 학습 시간을 필요로 하여 개발 지연과 한계가 발생한다. 따라서 HW 또는 SW 개선을 통해 파라미터 수, 학습 시간, 추론시간, 요구 메모리를 줄이는 연구가 지속 되어 왔다. 본 논문은 EfficientNet에서 사용된 Linear Bottleneck을 변경하여 정확도는 소폭 감소 하지만 기존 모델의 파라미터를 55%로 줄이는 경량화 모델을 제안한다.