• Journal of Internet Computing and Services
• /
• v.22 no.2
• /
• pp.59-68
• /
• 2021

As artificial intelligence(AI) technologies, which have made rapid growth recently, began to be applied to the marine environment such as ships, there have been active researches on the application of CNN-based models specialized for digital videos. In E-Navigation service, which is combined with various technologies to detect floating objects of clash risk to reduce human errors and prevent fires inside ships, real-time processing is of huge importance. More functions added, however, mean a need for high-performance processes, which raises prices and poses a cost burden on shipowners. This study thus set out to propose a method capable of processing information at a high rate while maintaining the accuracy by applying Quantization techniques of a deep learning model. First, videos were pre-processed fit for the detection of floating matters in the sea to ensure the efficient transmission of video data to the deep learning entry. Secondly, the quantization technique, one of lightweight techniques for a deep learning model, was applied to reduce the usage rate of memory and increase the processing speed. Finally, the proposed deep learning model to which video pre-processing and quantization were applied was applied to various embedded boards to measure its accuracy and processing speed and test its performance. The proposed method was able to reduce the usage of memory capacity four times and improve the processing speed about four to five times while maintaining the old accuracy of recognition.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.11a
• /
• pp.918-919
• /
• 2023

딥러닝 기술이 발전함에 따라 모델의 복잡성 역시 증가하고 있다. 본 연구에서는 모델 경량화를 위해 Knowledge Distillation 기법을 Crowd Counting Model에 적용했다. M-SFANet을 Teacher 모델로, 파라미터수가 적은 MCNN 모델을 Student 모델로 채택해 Knowledge Distillation을 적용한 결과, 기존의 MCNN 모델보다 성능을 향상했다. 이는 정확도와 메모리 효율성 측면에서 많은 개선을 이루어 컴퓨팅 리소스가 부족한 기기에서도 본 모델을 실행할 수 있어 많은 활용이 가능할 것이다.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2021.06a
• /
• pp.345-347
• /
• 2021

단일 이미지 초해상도는 딥러닝의 발전과 함께 놀라운 성능 향상이 이루어 졌다. 이러한 딥러닝 모델은 매우 많은 파라미터를 갖고 있어 많은 연산량과 메모리를 필요로 한다. 하지만 사용할 수 있는 리소스는 한정되어 있기 때문에 네트워크를 경량화 시키려는 연구도 지속되어 왔다. 본 논문에서는 Stochastic Weight Averaging (SWA) 알고리즘을 이용하여 상대적으로 적은 양의 메모리와 연산을 추가해 이미지 초해상도 모델의 성능을 높이고 안정적인 학습을 달성하였다. SWA 알고리즘을 적용한 모델은 그렇지 않은 모델에 비해 테스트셋에서 최대 0.13dB 의 성능 향상을 보였다.

• Journal of the Institute of Convergence Signal Processing
• /
• v.24 no.1
• /
• pp.21-26
• /
• 2023

Knowledge distillation (KD) is a model lightening technology that transfers the knowledge of deep models to light models. Most KD methods have been developed for classification models, and there have been few KD studies in the field of super-resolution (SR). In this paper, various KD methods are applied to an SR model and their performance is compared. Specifically, we modified the loss function to apply each KD method to the SR model and conducted an experiment to learn a student model that was about 27 times lighter than the teacher model and to double the image resolution. Through the experiment, it was confirmed that some KD methods were not valid when applied to SR models, and that the performance was the highest when the relational KD and the traditional KD methods were combined.

• Journal of the Korea Society of Computer and Information
• /
• v.26 no.10
• /
• pp.37-43
• /
• 2021

Using the acoustic features of speech, important social and linguistic information about the speaker can be obtained, and one of the key features is the dialect. A speaker's use of a dialect is a major barrier to interaction with a computer. Dialects can be distinguished at various levels such as phonemes, syllables, words, phrases, and sentences, but it is difficult to distinguish dialects by identifying them one by one. Therefore, in this paper, we propose a lightweight Korean dialect classification model using only MFCC among the features of speech data. We study the optimal method to utilize MFCC features through Korean conversational voice data, and compare the classification performance of five Korean dialects in Gyeonggi/Seoul, Gangwon, Chungcheong, Jeolla, and Gyeongsang in eight machine learning and deep learning classification models. The performance of most classification models was improved by normalizing the MFCC, and the accuracy was improved by 1.07% and F1-score by 2.04% compared to the best performance of the classification model before normalizing the MFCC.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.11a
• /
• pp.129-132
• /
• 2020

최근 모바일 기기에서 딥러닝 모델을 사용하기 위한 경량화 연구가 진행되고 있다. 그중 모델의 가중치 표현 bit를 줄이는 양자화와 사용하기 위한 다양한 압축 알고리즘이 개발되었다. 하지만 대부분의 양자화 및 압축 알고리즘들은 한 번 이상의 Fine-tuning을 거쳐야 하는데 이 과정은 모바일 환경에서 수행하기에는 연산복잡도가 너무 높다. 따라서 본 논문은 양자화된 가중치를 High Efficiency Video Coding(HEVC)을 통해 압축하는 방법을 제안하고 정확도와 압축률을 실험한다. 실험결과는 양자화만 실시한 경우 대비 크기는 25%의 감소했지만, 정확도는 0.7% 감소했다. 따라서 이런 결과는 모바일 기기에 가중치를 전송하는 과정에 적용될 수 있다.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2021.06a
• /
• pp.271-274
• /
• 2021

Fisheye 카메라로 촬영된 영상은 일반 영상보다 넓은 시야각을 갖는 장점으로 여러 분야에서 활용되고 있다. 그러나 fisheye 카메라로 촬영된 영상은 어안렌즈의 곡률로 인하여 영상의 중앙 부분은 팽창되고 외곽 부분은 축소되는 방사 왜곡이 발생하기 때문에 영상을 활용함에 있어서 어려움이 있다. 이러한 방사 왜곡을 보정하기 위하여 기존 영상처리 분야에서는 렌즈의 곡률을 수학적으로 계산하여 보정하기도 하지만 이는 각각의 렌즈마다 왜곡 파라미터를 추정해야 하기 때문에, 개별적인 GT (Ground Truth) 영상이 필요하다는 제한 사항이 있다. 이에 본 논문에서는 렌즈의 종류마다 GT 영상을 필요로 하는 기존 기술의 제한 사항을 극복하기 위하여, fisheye 영상만을 입력으로 하여 왜곡계수를 계산하는 딥러닝 네트워크를 제안하고자 한다. 또한, 단일 왜곡계수를 왜곡모델로 활용함으로써 layer 수를 크게 줄일 수 있는 경량화 네트워크를 제안한다.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.23 no.4
• /
• pp.41-46
• /
• 2023

Human urine is a process of excreting waste products in the blood, and it is easy to collect and contains various substances. Urinalysis is used to check for diseases, health conditions, and urinary tract infections. There are three methods of urinalysis: physical property test, chemical test, and microscopic test, and chemical test results can be easily confirmed using urine test strips. A variety of items can be tested on the urine test strip, through which various diseases can be identified. Recently, with the spread of smart phones, research on reading urine test strips using smart phones is being conducted. There is a method of detecting and reading the color change of a urine test strip using a smartphone. This method uses the RGB values and the color difference formula to discriminate. However, there is a problem in that accuracy is lowered due to various environmental factors. This paper applies a deep learning model to solve this problem. In particular, color discrimination of a urine test strip is improved in a smartphone using a lightweight CNN (Convolutional Neural Networks) model. CNN is a useful model for image recognition and pattern finding, and a lightweight version is also available. Through this, it is possible to operate a deep learning model on a smartphone and extract accurate urine test results. Urine test strips were taken in various environments to prepare deep learning model training images, and a urine test service application was designed using MobileNet V3.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.11a
• /
• pp.435-437
• /
• 2022

보행 보조 로봇의 효과적인 보조를 위해서는 사용자의 보행 유형을 인식하는 것이 중요하다. 본 논문에서는 end-to-end 분류 네트워크 기반 보행 환경 인식 방법을 사용하여 사용자의 보행 유형을 강인하게 추정한다. 실외 보행 환경을 오르막길, 평지, 내리막길 3 가지로 분류하는 딥러닝 모델을 학습시켰으며, 엣지 디바이스에서 이를 사용하기 위해 네트워크 경량화를 진행하였다. 경량화 후 추론 속도는 약 47FPS 수준으로 실시간으로 보행 보조 로봇에 적용 가능한 것을 검증했으며, 정확도 측면에서도 97% 이상의 성능을 얻을 수 있었다.

• Smart Media Journal
• /
• v.12 no.5
• /
• pp.28-35
• /
• 2023

In this study, we propose a new light-weight model RoutingConvNet with fewer parameters to improve the applicability and practicality of speech emotion recognition. To reduce the number of learnable parameters, the proposed model connects bidirectional MFCCs on a channel-by-channel basis to learn long-term emotion dependence and extract contextual features. A light-weight deep CNN is constructed for low-level feature extraction, and self-attention is used to obtain information about channel and spatial signals in speech signals. In addition, we apply dynamic routing to improve the accuracy and construct a model that is robust to feature variations. The proposed model shows parameter reduction and accuracy improvement in the overall experiments of speech emotion datasets (EMO-DB, RAVDESS, and IEMOCAP), achieving 87.86%, 83.44%, and 66.06% accuracy respectively with about 156,000 parameters. In this study, we proposed a metric to calculate the trade-off between the number of parameters and accuracy for performance evaluation against light-weight.