• Journal of the Korea Computer Graphics Society
• /
• v.29 no.5
• /
• pp.11-20
• /
• 2023

In this paper, we propose a novel deep learning-based motion reconstruction approach that facilitates the generation of full-body motions, including finger motions, while also enabling the online adjustment of motion generation delays. The proposed method combines the Vive Tracker with a deep learning method to achieve more accurate motion reconstruction while effectively mitigating foot skating issues through the use of an Inverse Kinematics (IK) solver. The proposed method utilizes a trained AutoEncoder to reconstruct character body motions using tracker data in real-time while offering the flexibility to adjust motion generation delays as needed. To generate hand motions suitable for the reconstructed body motion, we employ a Fully Connected Network (FCN). By combining the reconstructed body motion from the AutoEncoder with the hand motions generated by the FCN, we can generate full-body motions of characters that include hand movements. In order to alleviate foot skating issues in motions generated by deep learning-based methods, we use an IK solver. By setting the trackers located near the character's feet as end-effectors for the IK solver, our method precisely controls and corrects the character's foot movements, thereby enhancing the overall accuracy of the generated motions. Through experiments, we validate the accuracy of motion generation in the proposed deep learning-based motion reconstruction scheme, as well as the ability to adjust latency based on user input. Additionally, we assess the correction performance by comparing motions with the IK solver applied to those without it, focusing particularly on how it addresses the foot skating issue in the generated full-body motions.

• Journal of IKEEE
• /
• v.28 no.2
• /
• pp.150-157
• /
• 2024

In this paper, we propose a novel algorithm for predicting the number of apples on an apple tree using a deep learning-based object detection model and a polynomial regression model. Measuring the number of apples on an apple tree can be used to predict apple yield and to assess losses for determining agricultural disaster insurance payouts. To measure apple fruit load, we photographed the front and back sides of apple trees. We manually labeled the apples in the captured images to construct a dataset, which was then used to train a one-stage object detection CNN model. However, when apples on an apple tree are obscured by leaves, branches, or other parts of the tree, they may not be captured in images. Consequently, it becomes difficult for image recognition-based deep learning models to detect or infer the presence of these apples. To address this issue, we propose a two-stage inference process. In the first stage, we utilize an image-based deep learning model to count the number of apples in photos taken from both sides of the apple tree. In the second stage, we conduct a polynomial regression analysis, using the total apple count from the deep learning model as the independent variable, and the actual number of apples manually counted during an on-site visit to the orchard as the dependent variable. The performance evaluation of the two-stage inference system proposed in this paper showed an average accuracy of 90.98% in counting the number of apples on each apple tree. Therefore, the proposed method can significantly reduce the time and cost associated with manually counting apples. Furthermore, this approach has the potential to be widely adopted as a new foundational technology for fruit load estimation in related fields using deep learning.

• Korean Journal of Radiology
• /
• v.20 no.5
• /
• pp.749-758
• /
• 2019

Objective: To investigate whether a computer-aided diagnosis (CAD) system based on a deep learning framework (deep learning-based CAD) improves the diagnostic performance of radiologists in differentiating between malignant and benign masses on breast ultrasound (US). Materials and Methods: B-mode US images were prospectively obtained for 253 breast masses (173 benign, 80 malignant) in 226 consecutive patients. Breast mass US findings were retrospectively analyzed by deep learning-based CAD and four radiologists. In predicting malignancy, the CAD results were dichotomized (possibly benign vs. possibly malignant). The radiologists independently assessed Breast Imaging Reporting and Data System final assessments for two datasets (US images alone or with CAD). For each dataset, the radiologists' final assessments were classified as positive (category 4a or higher) and negative (category 3 or lower). The diagnostic performances of the radiologists for the two datasets (US alone vs. US with CAD) were compared Results: When the CAD results were added to the US images, the radiologists showed significant improvement in specificity (range of all radiologists for US alone vs. US with CAD: 72.8-92.5% vs. 82.1-93.1%; p < 0.001), accuracy (77.9-88.9% vs. 86.2-90.9%; p = 0.038), and positive predictive value (PPV) (60.2-83.3% vs. 70.4-85.2%; p = 0.001). However, there were no significant changes in sensitivity (81.3-88.8% vs. 86.3-95.0%; p = 0.120) and negative predictive value (91.4-93.5% vs. 92.9-97.3%; p = 0.259). Conclusion: Deep learning-based CAD could improve radiologists' diagnostic performance by increasing their specificity, accuracy, and PPV in differentiating between malignant and benign masses on breast US.

• Imaging Science in Dentistry
• /
• v.54 no.1
• /
• pp.81-91
• /
• 2024

Purpose: The objective of this study was to propose a deep-learning model for the detection of the mandibular canal on dental panoramic radiographs. Materials and Methods: A total of 2,100 panoramic radiographs (PANs) were collected from 3 different machines: RAYSCAN Alpha (n=700, PAN A), OP-100 (n=700, PAN B), and CS8100 (n=700, PAN C). Initially, an oral and maxillofacial radiologist coarsely annotated the mandibular canals. For deep learning analysis, convolutional neural networks (CNNs) utilizing U-Net architecture were employed for automated canal segmentation. Seven independent networks were trained using training sets representing all possible combinations of the 3 groups. These networks were then assessed using a hold-out test dataset. Results: Among the 7 networks evaluated, the network trained with all 3 available groups achieved an average precision of 90.6%, a recall of 87.4%, and a Dice similarity coefficient (DSC) of 88.9%. The 3 networks trained using each of the 3 possible 2-group combinations also demonstrated reliable performance for mandibular canal segmentation, as follows: 1) PAN A and B exhibited a mean DSC of 87.9%, 2) PAN A and C displayed a mean DSC of 87.8%, and 3) PAN B and C demonstrated a mean DSC of 88.4%. Conclusion: This multi-device study indicated that the examined CNN-based deep learning approach can achieve excellent canal segmentation performance, with a DSC exceeding 88%. Furthermore, the study highlighted the importance of considering the characteristics of panoramic radiographs when developing a robust deep-learning network, rather than depending solely on the size of the dataset.

• KIPS Transactions on Software and Data Engineering
• /
• v.11 no.3
• /
• pp.141-148
• /
• 2022

Recently, reinforcement learning combined with deep neural network technology has achieved remarkable success in various fields such as board games such as Go and chess, computer games such as Atari and StartCraft, and robot object manipulation tasks. However, such deep reinforcement learning describes states, actions, and policies in vector representation. Therefore, the existing deep reinforcement learning has some limitations in generality and interpretability of the learned policy, and it is difficult to effectively incorporate domain knowledge into policy learning. On the other hand, dNL-RRL, a new relational reinforcement learning framework proposed to solve these problems, uses a kind of vector representation for sensor input data and lower-level motion control as in the existing deep reinforcement learning. However, for states, actions, and learned policies, It uses a relational representation with logic predicates and rules. In this paper, we present dNL-RRL-based policy learning for transportation mobile robots in a manufacturing environment. In particular, this study proposes a effective method to utilize the prior domain knowledge of human experts to improve the efficiency of relational reinforcement learning. Through various experiments, we demonstrate the performance improvement of the relational reinforcement learning by using domain knowledge as proposed in this paper.

• Korean Journal of Remote Sensing
• /
• v.38 no.5_1
• /
• pp.511-521
• /
• 2022

Cloud detection and removal from satellite images is an essential process for topographic observation and analysis. Threshold-based cloud detection techniques show stable performance because they detect using the physical characteristics of clouds, but they have the disadvantage of requiring all channels' images and long computational time. Cloud detection techniques using deep learning, which have been studied recently, show short computational time and excellent performance even using only four or less channel (RGB, NIR) images. In this paper, we confirm the performance dependence of the deep learning network according to the heterogeneous learning dataset with different resolutions. The DeepLabV3+ network was improved so that channel features of cloud detection were extracted and learned with two published heterogeneous datasets and mixed data respectively. As a result of the experiment, clouds' Jaccard index was low in a network that learned with different kind of images from test images. However, clouds' Jaccard index was high in a network learned with mixed data that added some of the same kind of test data. Clouds are not structured in a shape, so reflecting channel features in learning is more effective in cloud detection than spatial features. It is necessary to learn channel features of each satellite sensors for cloud detection. Therefore, cloud detection of heterogeneous sensors with different resolutions is very dependent on the learning dataset.

• Journal of Electrical Engineering and Technology
• /
• v.12 no.1
• /
• pp.363-372
• /
• 2017

This paper presents a novel face recognition algorithm based on the deep convolution neural network and key point detection jointed local binary pattern methodology to enhance the accuracy of face recognition. We firstly propose the modified face key feature point location detection method to enhance the traditional localization algorithm to better pre-process the original face images. We put forward the grey information and the color information with combination of a composite model of local information. Then, we optimize the multi-layer network structure deep learning algorithm using the Fisher criterion as reference to adjust the network structure more accurately. Furthermore, we modify the local binary pattern texture description operator and combine it with the neural network to overcome drawbacks that deep neural network could not learn to face image and the local characteristics. Simulation results demonstrate that the proposed algorithm obtains stronger robustness and feasibility compared with the other state-of-the-art algorithms. The proposed algorithm also provides the novel paradigm for the application of deep learning in the field of face recognition which sets the milestone for further research.

• The Journal of Korea Robotics Society
• /
• v.11 no.3
• /
• pp.115-126
• /
• 2016

This paper suggests the method of the spherical signature description of 3D point clouds taken from the laser range scanner on the ground vehicle. Based on the spherical signature description of each point, the extractor of significant environmental features is learned by the Deep Belief Nets for the urban structure classification. Arbitrary point among the 3D point cloud can represents its signature in its sky surface by using several neighborhood points. The unit spherical surface centered on that point can be considered to accumulate the evidence of each angular tessellation. According to a kind of point area such as wall, ground, tree, car, and so on, the results of spherical signature description look so different each other. These data can be applied into the Deep Belief Nets, which is one of the Deep Neural Networks, for learning the environmental feature extractor. With this learned feature extractor, 3D points can be classified due to its urban structures well. Experimental results prove that the proposed method based on the spherical signature description and the Deep Belief Nets is suitable for the mobile robots in terms of the classification accuracy.

• Journal of Korea Multimedia Society
• /
• v.22 no.12
• /
• pp.1447-1456
• /
• 2019

As artificial intelligence technology is actively used in image processing, it is possible to generate high-quality fake images based on deep learning. Fake images generated using GAN(Generative Adversarial Network), one of unsupervised learning algorithms, have reached levels that are hard to discriminate from the naked eye. Detecting these fake images is required as they can be abused for crimes such as illegal content production, identity fraud and defamation. In this paper, we develop a deep-learning model based on CNN(Convolutional Neural Network) for the detection of StyleGAN fake images. StyleGAN is one of GAN algorithms and has an excellent performance in generating face images. We experiment with 48 number of experimental scenarios developed by combining parameters of the proposed model. We train and test each scenario with 300,000 number of real and fake face images in order to present a model parameter that improves performance in the detection of fake faces.

• The Journal of Information Systems
• /
• v.27 no.3
• /
• pp.123-140
• /
• 2018

Purpose The purpose of this study is to explore the optimal trading frequency which is useful for stock price prediction by using deep learning for charting image data. We also want to identify the appropriate time for accurate forecasting of stock price when performing pattern analysis. Design/methodology/approach In order to find the optimal trading frequency patterns and forecast timings, this study is performed as follows. First, stock price data is collected using OpenAPI provided by Daishin Securities, and candle chart images are created by data frequency and forecasting time. Second, the patterns are generated by the charting images and the learning is performed using the CNN. Finally, we find the optimal trading frequency patterns and forecasting timings. Findings According to the experiment results, this study confirmed that when the 10 minute frequency data is judged to be a decline pattern at previous 1 tick, the accuracy of predicting the market frequency pattern at which the market decreasing is 76%, which is determined by the optimal frequency pattern. In addition, we confirmed that forecasting of the sales frequency pattern at previous 1 tick shows higher accuracy than previous 2 tick and 3 tick.