• The Journal of Korea Robotics Society
• /
• v.18 no.1
• /
• pp.66-71
• /
• 2023

As cameras have become primary sensors for mobile robots, vision based Simultaneous Localization and Mapping (SLAM) has achieved impressive results with the recent development of computer vision and deep learning. However, vision information has a disadvantage in that a lot of information disappears in a low-light environment. To overcome the problem, we propose an image enhancement method to perform visual SLAM in a low-light environment. Using the deep generative adversarial models and modified gamma correction, the quality of low-light images were improved. The proposed method is less sharp than the existing method, but it can be applied to ORB-SLAM in real time by dramatically reducing the amount of computation. The experimental results were able to prove the validity of the proposed method by applying to public Dataset TUM and VIVID++.

• KIISE Transactions on Computing Practices
• /
• v.23 no.7
• /
• pp.417-423
• /
• 2017

Recently, many embedded devices that have the computing capability required for deep learning have become available; hence, many new applications using these devices are emerging. However, these embedded devices have an architecture different from that of PCs and high-performance servers. In this paper, we propose a method that improves the performance of deep-learning framework by considering the architecture of an embedded device that shares memory between the CPU and the GPU. The proposed method is implemented in Caffe, an open-source deep-learning framework, and is evaluated on an NVIDIA Jetson TK1 embedded device. In the experiment, we investigate the image recognition performance of several state-of-the-art deep-learning networks, including AlexNet, VGGNet, and GoogLeNet. Our results show that the proposed method can achieve significant performance gain. For instance, in AlexNet, we could reduce image recognition latency by about 33% and energy consumption by about 50%.

• Journal of Internet Computing and Services
• /
• v.18 no.5
• /
• pp.69-77
• /
• 2017

In this paper, a disaster response scheme based on deep learning and augmented reality technology is proposed and a network resource reservation scheme is presented accordingly. The features of deep learning, augmented reality technology and its relevance to the disaster areas are explained. Deep learning technology can be used to accurately recognize disaster situations and to implement related disaster information as augmented reality, and to enhance disaster response capabilities by providing disaster response On-site disaster response agent, ICS (Incident Command System) and MCS (Multi-agency Coordination Systems). In the case of various disasters, the fire situation is focused on and it is proposed that a plan to strengthen disaster response capability effectively by providing fire situation recognition based on deep learning and augmented reality information. Finally, a scheme to secure network resources to utilize the disaster response method of this paper is proposed.

• The Journal of the Acoustical Society of Korea
• /
• v.41 no.1
• /
• pp.30-37
• /
• 2022

This paper compares and evaluates model performance from two perspectives according to the learning target and network structure for training Deep Neural Network (DNN)-based speech enhancement models in the frequency domain. In this case, spectrum mapping and Time-Frequency (T-F) masking techniques were used as learning targets, and a real network and a complex network were used for the network structure. The performance of the speech enhancement model was evaluated through two objective evaluation metrics: Perceptual Evaluation of Speech Quality (PESQ) and Short-Time Objective Intelligibility (STOI) depending on the scale of the dataset. Test results show the appropriate size of the training data differs depending on the type of networks and the type of dataset. In addition, they show that, in some cases, using a real network may be a more realistic solution if the number of total parameters is considered because the real network shows relatively higher performance than the complex network depending on the size of the data and the learning target.

• Journal of Internet Computing and Services
• /
• v.24 no.4
• /
• pp.85-92
• /
• 2023

Deep learning applications are increasingly being leveraged for disease detection tasks in medical imaging modalities such as X-ray, Computed Tomography (CT), and Magnetic Resonance Imaging (MRI). Most data-centric deep learning challenges necessitate the use of supervised learning methodologies to attain high accuracy and to facilitate performance evaluation through comparison with the ground truth. Supervised learning mandates a substantial amount of image and label sets, however, procuring an adequate volume of medical imaging data for training is a formidable task. Various data augmentation strategies can mitigate the underfitting issue inherent in supervised learning-based models that are trained on limited medical image and label sets. This research investigates the enhancement of a deep learning-based rib fracture segmentation model and the efficacy of data augmentation techniques such as left-right flipping, rotation, and scaling. Augmented dataset with L/R flipping and rotations(30°, 60°) increased model performance, however, dataset with rotation(90°) and ⨯0.5 rescaling decreased model performance. This indicates the usage of appropriate data augmentation methods depending on datasets and tasks.

• Journal of Korean Tunnelling and Underground Space Association
• /
• v.21 no.3
• /
• pp.419-432
• /
• 2019

Most of deep learning model training was proceeded by supervised learning, which is to train labeling data composed by inputs and corresponding outputs. Labeling data was directly generated manually, so labeling accuracy of data is relatively high. However, it requires heavy efforts in securing data because of cost and time. Additionally, the main goal of supervised learning is to improve detection performance for 'True Positive' data but not to reduce occurrence of 'False Positive' data. In this paper, the occurrence of unpredictable 'False Positive' appears by trained modes with labeling data and 'True Positive' data in monitoring of deep learning-based CCTV accident detection system, which is under operation at a tunnel monitoring center. Those types of 'False Positive' to 'fire' or 'person' objects were frequently taking place for lights of working vehicle, reflecting sunlight at tunnel entrance, long black feature which occurs to the part of lane or car, etc. To solve this problem, a deep learning model was developed by simultaneously training the 'False Positive' data generated in the field and the labeling data. As a result, in comparison with the model that was trained only by the existing labeling data, the re-inference performance with respect to the labeling data was improved. In addition, re-inference of the 'False Positive' data shows that the number of 'False Positive' for the persons were more reduced in case of training model including many 'False Positive' data. By training of the 'False Positive' data, the capability of field application of the deep learning model was improved automatically.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.5
• /
• pp.439-451
• /
• 2021

We propose a deep learning-based beamformer combined with spectral subtraction for continuous speech recognition operating in noisy environments. Conventional beamforming systems were mostly evaluated by using pre-segmented audio signals which were typically generated by mixing speech and noise continuously on a computer. However, since speech utterances are sparsely uttered along the time axis in real environments, conventional beamforming systems degrade in case when noise-only signals without speech are input. To alleviate this drawback, we combine online beamforming algorithm and spectral subtraction. We construct a Continuous Speech Enhancement (CSE) evaluation set to evaluate the online beamforming algorithm in noisy environments. The evaluation set is built by mixing sparsely-occurring speech utterances of the CHiME3 evaluation set and continuously-played CHiME3 background noise and background music of MUSDB. Using a Kaldi-based toolkit and Google web speech recognizer as a speech recognition back-end, we confirm that the proposed online beamforming algorithm with spectral subtraction shows better performance than the baseline online algorithm.

• Journal of the Korea Convergence Society
• /
• v.11 no.12
• /
• pp.15-22
• /
• 2020

In this paper, we proposed a deep learning based super-resolution method that combines Channel Attention and Spatial Attention feature enhancement methods. It is important to restore high-frequency components, such as texture and features, that have large changes in surrounding pixels during super-resolution processing. We proposed a super-resolution method using feature enhancement that combines Channel Attention and Spatial Attention. The existing CNN (Convolutional Neural Network) based super-resolution method has difficulty in deep network learning and lacks emphasis on high frequency components, resulting in blurry contours and distortion. In order to solve the problem, we used an emphasis block that combines Channel Attention and Spatial Attention to which Skip Connection was applied, and a Residual Block. The emphasized feature map extracted by the method was extended through Sub-pixel Convolution to obtain the super resolution. As a result, about PSNR improved by 5%, SSIM improved by 3% compared with the conventional SRCNN, and by comparison with VDSR, about PSNR improved by 2% and SSIM improved by 1%.

• Journal of Internet of Things and Convergence
• /
• v.8 no.2
• /
• pp.79-84
• /
• 2022

This paper presents an enhancement method of document restoration capability which is robust for security, loss, and contamination, It is based on two methods, that is, encryption and DnCNN(DeNoise Convolution Neural Network). In order to implement this encryption method, a mathematical model is applied as a spatial frequency transfer function used in optics of 2D image information. Then a method is proposed with optical interference patterns as encryption using spatial frequency transfer functions and using mathematical variables of spatial frequency transfer functions as ciphers. In addition, by applying the DnCNN method which is bsed on deep learning technique, the restoration capability is enhanced by removing noise. With an experimental evaluation, with 65% information loss, by applying Pre-Training DnCNN Deep Learning, the peak signal-to-noise ratio (PSNR) shows 11% or more superior in compared to that of the spatial frequency transfer function only. In addition, it is confirmed that the characteristic of CC(Correlation Coefficient) is enhanced by 16% or more.

• International Journal of Computer Science & Network Security
• /
• v.24 no.2
• /
• pp.158-168
• /
• 2024

There is hardly any person in modern times who has not taken soft drinks instead of drinking water. The rate of people taking soft drinks being surprisingly high, researchers around the world have cautioned from time to time that these drinks lead to weight gain, raise the risk of non-communicable diseases and so on. Therefore, in this work an image-based tool is developed to monitor the nutritional information of soft drinks by using deep convolutional neural network with transfer learning. At first, visual saliency, mean shift segmentation, thresholding and noise reduction technique, collectively known as 'pre-processing' are adopted to extract the location of drinks region. After removing backgrounds and segment out only the desired area from image, we impose Discrete Wavelength Transform (DWT) based resolution enhancement technique is applied to improve the quality of image. After that, transfer learning model is employed for the classification of drinks. Finally, nutrition value of each drink is estimated using Bag-of-Feature (BoF) based classification and Euclidean distance-based ratio calculation technique. To achieve this, a dataset is built with ten most consumed soft drinks in Bangladesh. These images were collected from imageNet dataset as well as internet and proposed method confirms that it has the ability to detect and recognize different types of drinks with an accuracy of 98.51%.