Noh, Hyung-nam. 2017. “Entertainment Science Based on Deep Learning: focused on Areal Sociolinguistics”. The Sociolinguistic Journal of Korea 25(1). 27~52. The aim of this paper is to suggest a new scientific discipline in sociolinguistic research, dealing with entertainment science based on deep learning focused on areal sociolinguistics as a current methodology de facto made by ultra-fusion of area studies and sociolinguistics. From a fact-oriented and data-oriented analysis perspective this paper examines real phenomena of areal sociolinguistics provoked by two famous sing-a-song writers: America’s Robert Allen Zimmerman, so-called 2016 Nobel prize winner Bob Dylan, and Brazil’s Paulo Coelho de Souza. The results of the qualitative analyses between two eminent areas, where particular attributes of alternative societies are filled with swarm intelligence on the basis of resistance consciousness, suggest the areal sociolinguistics mentioned-above. From the diachronic and synchronic viewpoints of cross-over geographical cultures this paper makes a mid-range generalization, on making a definition about alternative societies in America and Brazil in spite of the geographical methodology of area studies between the two countries, being offered by stubborn resistance against ready-made ideas to calm down keen psychological conflicts among established moral principles to overcome philosophical catastrophe in social chaos, and full of competitive instinct against existing generations.

In this paper, we present a learning platform for robotic grasping in real world, in which actor-critic deep reinforcement learning is employed to directly learn the grasping skill from raw image pixels and rarely observed rewards. This is a challenging task because existing algorithms based on deep reinforcement learning require an extensive number of training data or massive computational cost so that they cannot be affordable in real world settings. To address this problems, the proposed learning platform basically consists of two training phases; a learning phase in simulator and subsequent learning in real world. Here, main processing blocks in the platform are extraction of latent vector based on state representation learning and disentanglement of a raw image, generation of adapted synthetic image using generative adversarial networks, and object detection and arm segmentation for the disentanglement. We demonstrate the effectiveness of this approach in a real environment.

This paper proposes a model and train method that can real-time detect objects and distances estimation based on a monocular camera by applying deep learning. It used YOLOv2 model which is applied to autonomous or robot due to the fast image processing speed. We have changed and learned the loss function so that the YOLOv2 model can detect objects and distances at the same time. The YOLOv2 loss function added a term for learning bounding box values x, y, w, h, and distance values z as 클래스ification losses. In addition, the learning was carried out by multiplying the distance term with parameters for the balance of learning. we trained the model location, recognition by camera and distance data measured by lidar so that we enable the model to estimate distance and objects from a monocular camera, even when the vehicle is going up or down hill. To evaluate the performance of object detection and distance estimation, MAP (Mean Average Precision) and Adjust R square were used and performance was compared with previous research papers. In addition, we compared the original YOLOv2 model FPS (Frame Per Second) for speed measurement with FPS of our model.

본 연구에서는 신속하고 편리한 고속도로 유지관리를 위해 주행 중인 차량에서 실시간으로 영상을 수집하고 분석하여 현장의 노면 포장 상태를 모니터링하고 손상을 탐지하는 딥러닝 기반의 고속도로 포장 손상 조사 기술에 대해 소개한다.

딥러닝 모델은 주어진 학습용 데이터에서 탐지하고자 하는 물체의 특징을 추출하기 때문에, 딥러닝 모델 학습을 위한 학습용 데이터 구축은 매우 중요하다. 본 연구에서는 균열을 탐지하는 딥러닝 모델의 성능을 향상시키기 위해, 실제 콘크리트 구조물이나 아스팔트 도로 표면에서 자주 발견될 수 있는 나뭇가지, 거미줄, 전선 등을 학습 데이터에 자동으로 포함시키고, negative 영역으로 분류하는 알고리즘을 개발하였다. 제안된 알고리즘을 사용하여 학습된 딥러닝 모델을 실제 도로 표면에 발생한 균열 탐지에 적용하여 실제 균열 탐지에 사용될 수 있음을 보였다.

최근 사회기반시설(SOC)의 증가와 노후화에 따라 기존의 인력중심의 육안검사를 기반으로 한 안전점검은 경제성과 안전성, 효율성 면에서 한계를 가지고 있다. 본 연구에서는 육안점검의 한계를 개선하기 위해 딥러닝 모델 기반 물체를 탐지하는 기술을 활용하여 터널 콘크리트 균열을 자동으로 탐지하는 기술을 개발하였으며, 이를 실제 터널 영상에 적용하여 그 성능을 검증하였다.

This paper presents a new benchmark system for visual odometry (VO) and monocular depth estimation (MDE). As deep learning has become a key technology in computer vision, many researchers are trying to apply deep learning to VO and MDE. Just a couple of years ago, they were independently studied in a supervised way, but now they are coupled and trained together in an unsupervised way. However, before designing fancy models and losses, we have to customize datasets to use them for training and testing. After training, the model has to be compared with the existing models, which is also a huge burden. The benchmark provides input dataset ready-to-use for VO and MDE research in ‘tfrecords’ format and output dataset that includes model checkpoints and inference results of the existing models. It also provides various tools for data formatting, training, and evaluation. In the experiments, the exsiting models were evaluated to verify their performances presented in the corresponding papers and we found that the evaluation result is inferior to the presented performances.

Carbonation of reinforced concrete is a major factor in the deterioration of reinforced concrete, and prediction of the resistance to carbonation is important in determining the durability life of reinforced concrete structures. In this study, basic research on the prediction of carbonation penetration depth of concrete using Deep Learning algorithm among artificial neural network theory was carried out. The data used in the experiment were analyzed by deep running algorithm by setting W/B, cement and blast furnace slag, fly ash content, relative humidity of the carbonated laboratory, temperature, CO2 concentration, Deep learning algorithms were used to study 60,000 times, and the analysis of the number of hidden layers was compared.

Last few years, many researches on deep learning-based crack detection model have been reported in order to develop an efficient structure inspection method. While developing crack detection deep learning model, many research results reported the importance of the training data. Since most of the research results only qualitatively discussed the importance of training data, this study examine the influence of the training data by experiment, especially in the case of negative samples such as construction joint, spider web and concrete blocks.

Construction safety is one of the significant problems on the world. Deep learning is an emerging term that acquires, processes and analyses image or video data to help computers have a high-level visual understanding of the world. In recent years, it has been introduced into the construction industry for improvements of occupational health and safety. This research contributes in solving this problem by using deep learning only RGB images that output detects the hazard zone on construction sites. The main goal of this study is to use different computer vision and deep learning to develop for different cases concerning fall related hazards.

The damage detection method of blade systems largely depends on the personal ability of an inspector using a camera. Thus, this paper proposes a deep learning-based detection method that can rapidly and reliably identify and evaluate the damages on the blades.

This paper proposes a deep learning-based underground object classification technique incorporated with phase analysis of ground penetrating radar (GPR) for enhancing the underground object classification capability. Deep convolutional neural network (CNN) using the combination of the B- and C-scan images has recently emerged for automated underground object classification. However, it often leads to misclassification because arbitrary underground objects may have similar signal features. To overcome the drawback, the combination of B- and C-scan images as well as phase information of GPR are simultaneously used for CNN in this study, enabling to have more distinguishable signal features among various underground objects. The proposed technique is validated using in-situ GPR data obtained from urban roads in Seoul, South Korea. The validation results show that the false alarm is significantly reduced compared to the CNN results using only B- and C-scan images.

Pavement condition deteriorates due to various environmental issues. This can be seen on the pavement surface as a form of distress. A crack can be considered as a typical form of pavement distress in which it may reveal a critical condition of the road. Therefore, automatic and accurate detection of pavement crack and segmentation are crucial for pavement condition assessment and maintenance.

For effective human-robot interaction, robots need to understand the current situation context well, but also the robots need to transfer its understanding to the human participant in efficient way. The most convenient way to deliver robot’s understanding to the human participant is that the robot expresses its understanding using voice and natural language. Recently, the artificial intelligence for video understanding and natural language process has been developed very rapidly especially based on deep learning. Thus, this paper proposes robot vision to audio description method using deep learning. The applied deep learning model is a pipeline of two deep learning models for generating natural language sentence from robot vision and generating voice from the generated natural language sentence. Also, we conduct the real robot experiment to show the effectiveness of our method in human-robot interaction.

In the ground environment, mobile robot research uses sensors such as GPS and optical cameras to localize surrounding landmarks and to estimate the position of the robot. However, an underwater environment restricts the use of sensors such as optical cameras and GPS. Also, unlike the ground environment, it is difficult to make a continuous observation of landmarks for location estimation. So, in underwater research, artificial markers are installed to generate a strong and lasting landmark. When artificial markers are acquired with an underwater sonar sensor, different types of noise are caused in the underwater sonar image. This noise is one of the factors that reduces object detection performance. This paper aims to improve object detection performance through distortion and rotation augmentation of training data. Object detection is detected using a Faster R-CNN.

Recently, smart factories have attracted much attention as a result of the 4th Industrial Revolution. Existing factory automation technologies are generally designed for simple repetition without using vision sensors. Even small object assemblies are still dependent on manual work. To satisfy the needs for replacing the existing system with new technology such as bin picking and visual servoing, precision and real-time application should be core. Therefore in our work we focused on the core elements by using deep learning algorithm to detect and classify the target object for real-time and analyzing the object features. We chose YOLO CNN which is capable of real-time working and combining the two tasks as mentioned above though there are lots of good deep learning algorithms such as Mask R-CNN and Fast R-CNN. Then through the line and inside features extracted from target object, we can obtain final outline and estimate object posture.

본 연구는 물리적 수리·수문모형의 적용이 제한적인 감조하천에서의 수위예측을 목적으로 하고 있으며, 이를 위해 한강 잠수교를 대상으로 딥러닝 오픈소스 소프트웨어 라이브러리인 TensorFlow를 활용하여 LSTM 모형을 구성하고 2011년부터 2017년까지의 10분 단위의 잠수교 수위, 팔당 댐 방류량과 한강하구 강화대교지점의 예측조위 자료를 이용하여 모형학습(2011~2016) 및 수위예측(2017)을 수행하였다. 모형 매개변수는 민감도 분석을 통해 은닉층의 개수는 6개, 학습속도는 0.01, 학습횟수는 3000번로 결정하였으며, 모형 학습 시 학습정보의 시간적 양을 결정하는 중요한 매개변수인 시퀀스길이는 1시간, 3시간, 6시간으로 변화시키며 모의하였다. 최종적으로 선행시간에 따른 모의 예측능력을 평가하기 위해 LSTM 모형의 예측 선행시간을 6개(1 ~ 24시간)로 구분하여 실측수위와 예측수위와의 비교·분석을 수행한 결과, LSTM 모형의 최적의 성능을 내 는 결과는 시퀀스길이를 1시간으로 하였을 때로 분석되었으며, 특히 선행시간 1시간에 대한 예측정확도는 RMSE는 0.065 m, NSE는 0.99로 실 측수위에 매우 근접한 예측 결과를 나타내었다. 또한 시퀀스길이에 상관없이 선행시간이 길어질수록 모형의 예측 정확도는 2017년 전기간에 걸쳐 평균적으로 RMSE 0.08 m에서 0.28 m로 오차가 증가하였으며, NSE는 0.99에서 0.74로 감소하였다.

최근 다수의 분야에서 딥 러닝을 통한 연구 성과들이 사람의 판단력에 근접하는 결과를 보여주고 있다. 그리고 게임 산업에서는 온라인 커뮤니티, SNS의 활성화가 게임 흥행 여부를 결정할 정도로 중요성이 높아지고 있다. 본 연구는 딥 러닝을 이용해 온라인 커뮤니티, SNS에서 활동할 수 있는 시스템을 구성하고, 온라인 공간에서 사람들이 작성한 텍스트를 읽고 그에 대한 반응을 생성하고 스케쥴에 따라 트위터에 올리는 것을 목표로 한다. 순환 신경망(Recurrent Neural Network)을 이용해 텍스트를 생성하고 글 작성 스케쥴을 생성하는 모델들을 구성했고, 생성한 시각에 맞춰 모델들에 뉴스 제목을 입력해 댓글을 출력 받고 트위터에 작성하는 프로그램을 구현했다. 본 연구 결과는 온라인 게임 커뮤니티 활성화, Q&A 서비스 등에 적용이 가능할 것으로 예상된다.

영상 인식 기술은 평면 영상에 대해서 많이 연구되고 그 성능 또한 발전하고 있다. 그러나 평면 영상이 아닌 구면 파노라마 영상과 다양한 환경에서 주어지는 특수한 형태의 영상에 대한 인식은 평면과 다르게 기하학적인 왜곡으로 인해서 많은 어려움이 따른다. 본 논문에서는 평면영상의 인식 기술에서 최근 각광받는 훈련을 통한 신경망 인식 기법이 구면 파노라마 영상의 인식에서도 쓰일 수 있음을 보인다. 또한 구면 영상에 대한 기존 신경망 모델의 인식률을 높이기 위해서 큐브맵 변환을 활용하는 방법을 제시한다.