• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2016.06a
• /
• pp.37-39
• /
• 2016

최근 Convolutional neural networks(CNN) 기반의 초해상화 기법인 Super-Resolution Convolutional Neural Networks (SRCNN) 이 좋은 PSNR 성능을 발휘하는 것으로 보고되었다 [1]. 하지만 많은 제안 방법들이 고주파 성분을 복원하는데 한계를 드러내는 것처럼, SRCNN 도 고주파 성분 복원에 한계점을 지니고 있다. 또한 SRCNN 의 네트워크 층을 깊게 만들면 좋은 PSNR 성능을 발휘하는 것으로 널리 알려져 있지만, 네트워크의 층을 깊게 하는 것은 네트워크 파라미터 학습을 어렵게 하는 경향이 있다. 네트워크의 층을 깊게 할 경우, gradient 값이 아래(역방향) 층으로 갈수록 발산하거나 0 으로 수렴하여, 네트워크 파라미터 학습이 제대로 되지 않는 현상이 발생하기 때문이다. 따라서 본 논문에서는 네트워크 층을 깊게 하는 대신에, 입력을 다중 채널로 구성하여, 네트워크에 고주파 성분에 관한 추가적인 정보를 주는 방법을 제안하였다. 많은 초해상화 기법들이 고주파 성분의 복원 능력이 부족하다는 점에 착안하여, 우리는 네트워크가 고주파 성분에 관한 많은 정보를 필요로 한다는 것을 가정하였다. 따라서 우리는 네트워크의 입력을 고주파 성분이 여러 가지 강도로 입력되도록 저해상도 입력 영상들을 구성하였다. 또한 잔차신호 네트워크(residual networks)를 도입하여, 네트워크 파라미터를 학습할 때 고주파 성분의 복원에 집중할 수 있도록 하였다. 본 논문의 효율성을 검증하기 위하여 set5 데이터와 set14 데이터에 관하여 실험을 진행하였고, SRCNN 과 비교하여 set5 데이터에서는 2, 3, 4 배에 관하여 각각 평균 0.29, 0.35, 0.17dB 의 PSNR 성능 향상이 있었으며, set14 데이터에서는 3 배의 관하여 평균 0.20dB 의 PSNR 성능 향상이 있었다.

• Journal of Soil and Groundwater Environment
• /
• v.25 no.3
• /
• pp.74-83
• /
• 2020

Since fine dust (PM₁₀) has a significant influence on soil and groundwater composition during dry and wet deposition processes, it is of a vital importance to understand the fate and transport of aerosol in geological environments. Fine dust is formed after the chemical reaction of several precursors, typically observed in short intervals within a few hours. In this study, deep learning approach was applied to predict the fate of fine dust in an urban area. Deep learning training was performed by combining convolutional neural network (CNN) and recurrent neural network (RNN) techniques. The PM₁₀ concentration after 1 hour was predicted based on three-hour data by setting SO₂, CO, O₃, NO₂, and PM₁₀ as training data. The obtained coefficient of determination value, R², was 0.8973 between predicted and measured values for the entire concentration range of PM₁₀, suggesting deep learning method can be developed into a reliable and viable tool for prediction of fine dust concentration.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.19 no.3
• /
• pp.693-698
• /
• 2018

This paper proposes a vehicle detection method for parking areas using unmanned aerial vehicles (UAVs) and using YOLOv2, which is a recent, known, fast, object-detection real-time algorithm. The YOLOv2 convolutional network algorithm can calculate the probability of each class in an entire image with a one-pass evaluation, and can also predict the location of bounding boxes. It has the advantage of very fast, easy, and optimized-at-detection performance, because the object detection process has a single network. The sliding windows methods and region-based convolutional neural network series detection algorithms use a lot of region proposals and take too much calculation time for each class. So these algorithms have a disadvantage in real-time applications. This research uses the YOLOv2 algorithm to overcome the disadvantage that previous algorithms have in real-time processing problems. Using Darknet, OpenCV, and the Compute Unified Device Architecture as open sources for object detection. a deep learning server is used for the learning and detecting process with each car. In the experiment results, the algorithm could detect cars in a dense area using UAVs, and reduced overhead for object detection. It could be applied in real time.

• Korean Journal of Artificial Intelligence
• /
• v.5 no.1
• /
• pp.10-17
• /
• 2017

This study is the first attempt to raise face recognition ability through machine learning algorithm and apply to CRM's information gathering, analysis and application. In other words, through face recognition of VIP customer in distribution field, we can proceed more prompt and subdivided customized services. The interest in machine learning, which is used to implement artificial intelligence, has increased, and it has become an age to automate it by using machine learning beyond the way that a person directly models an object recognition process. Among them, Deep Learning is evaluated as an advanced technology that shows amazing performance in various fields, and is applied to various fields of image recognition. Face recognition, which is widely used in real life, has been developed to recognize criminals' faces and catch criminals. In this study, two image analysis models, TF-SLIM and Inception-V3, which are likely to be used for criminal face recognition, were selected, analyzed, and implemented. As an evaluation criterion, the image recognition model was evaluated based on the accuracy of the face recognition program which is already being commercialized. In this experiment, it was evaluated that the recognition accuracy was good when the accuracy of the image classification was more than 90%. A limit of our study which is a way to raise face recognition is left as a further research subjects.

• Journal of Broadcast Engineering
• /
• v.23 no.3
• /
• pp.361-368
• /
• 2018

As autonomous driving research is rapidly developing, pedestrian detection study is also successfully investigated. However, most of the study utilizes color image datasets and those are relatively easy to detect the pedestrian. In case of color images, the scene should be exposed by enough light in order to capture the pedestrian and it is not easy for the conventional methods to detect the pedestrian if it is the other case. Therefore, in this paper, we propose deep neural network (DNN)-based multi-spectrum pedestrian detection method using color and thermal images. Based on single-shot multibox detector (SSD), we propose fusion network structures which simultaneously employ color and thermal images. In the experiment, we used KAIST dataset. We showed that proposed SSD-H (SSD-Halfway fusion) technique shows 18.18% lower miss rate compared to the KAIST pedestrian detection baseline. In addition, the proposed method shows at least 2.1% lower miss rate compared to the conventional halfway fusion method.

• Journal of Broadcast Engineering
• /
• v.23 no.5
• /
• pp.642-655
• /
• 2018

Occupancy sensors installed in buildings and households turn off the light if the space is vacant. Currently PIR(pyroelectric infra-red) motion sensors have been utilized. Recently, the researches using camera sensors have been carried out in order to overcome the demerit of PIR that cannot detect stationary people. The detection of moving and stationary people is a main functionality of the occupancy sensors. In this paper, we propose an on-line human occupancy tracking method using convolutional neural network (CNN) and self-organizing map. It is well known that a large number of training samples are needed to train the model offline. To solve this problem, we use an untrained model and update the model by collecting training samples online directly from the test sequences. Using videos capurted from an overhead camera, experiments have validated that the proposed method effectively tracks human.

• Journal of Korea Society of Industrial Information Systems
• /
• v.25 no.3
• /
• pp.21-30
• /
• 2020

With the widespread of smartphones, the Quick response (QR) code became one of the most popular codes. In this paper, a new type of QR code is proposed to increase the storage capacities and also to contain private information by changing the colors and the shape of patterns in the codes. Then, for a variety of applications of the printed QR codes, this paper proposes an efficient authentic certification system, which is built on an conventional CNN (Convolutional neural network) architecture - VGGNet and classifies authentic or counterfeit with smartphones, easily. For authentic codes, the proposed system extracts the embedded private information. Through practical experiments with a printed QR code, it is shown that the proposed system can classify authentic or counterfeit code, perfectly, and also, are useful for extracting private information.

• Journal of Advanced Navigation Technology
• /
• v.24 no.2
• /
• pp.155-162
• /
• 2020

Recently, with the development of computer vision and deep learning technology, research on human action recognition has been actively conducted for video analysis, video surveillance, interactive multimedia, and human machine interaction applications. Diverse techniques have been introduced for human action understanding and classification by many researchers using RGB image, depth image, skeleton and inertial data. However, skeleton-based action discrimination is still a challenging research topic for human machine-interaction. In this paper, we propose an end-to-end skeleton joints mapping of action for generating spatio-temporal image so-called dynamic image. Then, an efficient deep convolution neural network is devised to perform the classification among the action classes. We use publicly accessible UTD-MHAD skeleton dataset for evaluating the performance of the proposed method. As a result of the experiment, the proposed system shows better performance than the existing methods with high accuracy of 97.45%.

• Journal of Practical Engineering Education
• /
• v.11 no.2
• /
• pp.175-180
• /
• 2019

Existing fashion web sites show only the search results for one type of clothes in items such as tops and bottoms. As the fashion market grows, consumers are demanding a platform to find a variety of fashion information. To solve this problem, we devised the idea of linking image classification through deep learning with a website and integrating SNS functions. User uploads their own image to the web site and uses the deep learning server to identify, classify and store the image's characteristics. Users can use the stored information to search for the images in various combinations. In addition, communication between users can be actively performed through the SNS function. Through this, the plan to solve the problem of existing fashion-related sites was prepared.

• Journal of KIISE
• /
• v.44 no.12
• /
• pp.1275-1281
• /
• 2017

The supervised learning approach is suitable for classification of insulting sentences, but pre-decided training sentences are necessary. Since a Character-level Convolution Neural Network is robust for each character, so is appropriate for classifying abusive sentences, however, has a drawback that demanding a lot of training sentences. In this paper, we propose transfer learning method that reusing the trained filters in the real classification process after the filters get the characteristics of offensive words by generated abusive/normal pair of sentences. We got higher performances of the classifier by decreasing the effects of data shortage and class imbalance. We executed experiments and evaluations for three datasets and got higher F1-score of character-level CNN classifier when applying transfer learning in all datasets.