• Korean Journal of Radiology
• /
• v.20 no.8
• /
• pp.1275-1284
• /
• 2019

Objective: To develop algorithms using convolutional neural networks (CNNs) for automatic segmentation of acute ischemic lesions on diffusion-weighted imaging (DWI) and compare them with conventional algorithms, including a thresholding-based segmentation. Materials and Methods: Between September 2005 and August 2015, 429 patients presenting with acute cerebral ischemia (training:validation:test set = 246:89:94) were retrospectively enrolled in this study, which was performed under Institutional Review Board approval. Ground truth segmentations for acute ischemic lesions on DWI were manually drawn under the consensus of two expert radiologists. CNN algorithms were developed using two-dimensional U-Net with squeeze-and-excitation blocks (U-Net) and a DenseNet with squeeze-and-excitation blocks (DenseNet) with squeeze-and-excitation operations for automatic segmentation of acute ischemic lesions on DWI. The CNN algorithms were compared with conventional algorithms based on DWI and the apparent diffusion coefficient (ADC) signal intensity. The performances of the algorithms were assessed using the Dice index with 5-fold cross-validation. The Dice indices were analyzed according to infarct volumes (< 10 mL, ≥ 10 mL), number of infarcts (≤ 5, 6-10, ≥ 11), and b-value of 1000 (b1000) signal intensities (< 50, 50-100, > 100), time intervals to DWI, and DWI protocols. Results: The CNN algorithms were significantly superior to conventional algorithms (p < 0.001). Dice indices for the CNN algorithms were 0.85 for U-Net and DenseNet and 0.86 for an ensemble of U-Net and DenseNet, while the indices were 0.58 for ADC-b1000 and b1000-ADC and 0.52 for the commercial ADC algorithm. The Dice indices for small and large lesions, respectively, were 0.81 and 0.88 with U-Net, 0.80 and 0.88 with DenseNet, and 0.82 and 0.89 with the ensemble of U-Net and DenseNet. The CNN algorithms showed significant differences in Dice indices according to infarct volumes (p < 0.001). Conclusion: The CNN algorithm for automatic segmentation of acute ischemic lesions on DWI achieved Dice indices greater than or equal to 0.85 and showed superior performance to conventional algorithms.

• KIPS Transactions on Software and Data Engineering
• /
• v.8 no.3
• /
• pp.115-122
• /
• 2019

For Korean phoneme recognition, Hidden Markov-Gaussian Mixture model(HMM-GMM) or hybrid models which combine artificial neural network with HMM have been mainly used. However, current approach has limitations in that such models require force-aligned corpus training data that is manually annotated by experts. Recently, researchers used neural network based phoneme recognition model which combines recurrent neural network(RNN)-based structure with connectionist temporal classification(CTC) algorithm to overcome the problem of obtaining manually annotated training data. Yet, in terms of implementation, these RNN-based models have another difficulty in that the amount of data gets larger as the structure gets more sophisticated. This problem of large data size is particularly problematic in the Korean language, which lacks refined corpora. In this study, we introduce CTC algorithm that does not require force-alignment to create a Korean phoneme recognition model. Specifically, the phoneme recognition model is based on convolutional neural network(CNN) which requires relatively small amount of data and can be trained faster when compared to RNN based models. We present the results from two different experiments and a resulting best performing phoneme recognition model which distinguishes 49 Korean phonemes. The best performing phoneme recognition model combines CNN with 3hop Bidirectional LSTM with the final Phoneme Error Rate(PER) at 3.26. The PER is a considerable improvement compared to existing Korean phoneme recognition models that report PER ranging from 10 to 12.

• Journal of Korea Society of Industrial Information Systems
• /
• v.27 no.6
• /
• pp.13-23
• /
• 2022

In this paper, we propose a detailed component image classification algorithm by fashion item for unstructured data retrieval in the fashion field. Due to the COVID-19 environment, AI-based online shopping malls are increasing recently. However, there is a limit to accurate unstructured data search with existing keyword search and personalized style recommendations based on user surfing behavior. In this study, pre-processing using Mask R-CNN was conducted using images crawled from online shopping sites and then classified components for each fashion item through CNN. We obtain the accuaracy for collar of the shirt's as 93.28%, the pattern of the shirt as 98.10%, the 3 classese fit of the jeans as 91.73%, And, we further obtained one for the 4 classes fit of jeans as 81.59% and the color of the jeans as 93.91%. At the results for the decorated items, we also obtained the accuract of the washing of the jeans as 91.20% and the demage of jeans accuaracy as 92.96%.

• Nuclear Engineering and Technology
• /
• v.55 no.8
• /
• pp.3030-3038
• /
• 2023

Nowadays, mobile robots have been used to search for uncontrolled radioactive source in indoor environments to avoid radiation exposure for technicians. However, in the indoor environments, especially in the presence of obstacles, how to make the robots with limited sensing capabilities automatically search for the radioactive source remains a major challenge. Also, the source search efficiency of robots needs to be further improved to meet practical scenarios such as limited exploration time. This paper proposes an automatic source search strategy, abbreviated as ACA: the location of source is estimated by a convolutional neural network (CNN), and the path is planned by the A-star algorithm. First, the search area is represented as an occupancy grid map. Then, the radiation dose distribution of the radioactive source in the occupancy grid map is obtained by Monte Carlo (MC) method simulation, and multiple sets of radiation data are collected through the eight neighborhood self-avoiding random walk (ENSAW) algorithm as the radiation data set. Further, the radiation data set is fed into the designed CNN architecture to train the network model in advance. When the searcher enters the search area where the radioactive source exists, the location of source is estimated by the network model and the search path is planned by the A-star algorithm, and this process is iterated continuously until the searcher reaches the location of radioactive source. The experimental results show that the average number of radiometric measurements and the average number of moving steps of the ACA algorithm are only 2.1% and 33.2% of those of the gradient search (GS) algorithm in the indoor environment without obstacles. In the indoor environment shielded by concrete walls, the GS algorithm fails to search for the source, while the ACA algorithm successfully searches for the source with fewer moving steps and sparse radiometric data.

• Korean Journal of Remote Sensing
• /
• v.37 no.5_3
• /
• pp.1475-1490
• /
• 2021

Detecting oil spill area using statistical characteristics of SAR images has limitations in that classification algorithm is complicated and is greatly affected by outliers. To overcome these limitations, studies using neural networks to classify oil spills are recently investigated. However, the studies to evaluate whether the performance of model shows a consistent detection performance for various oil spill cases were insufficient. Therefore, in this study, two CNNs (Convolutional Neural Networks) with basic structures(Simple CNN and U-net) were used to discover whether there is a difference in detection performance according to the structure of CNN and distribution characteristics of oil spill. As a result, through the method proposed in this study, the Simple CNN with contracting path only detected oil spill with an F1 score of 86.24% and U-net, which has both contracting and expansive path showed an F1 score of 91.44%. Both models successfully detected oil spills, but detection performance of the U-net was higher than Simple CNN. Additionally, in order to compare the accuracy of models according to various oil spill cases, the cases were classified into four different categories according to the spatial distribution characteristics of the oil spill (presence of land near the oil spill area) and the clarity of border between oil and seawater. The Simple CNN had F1 score values of 85.71%, 87.43%, 86.50%, and 85.86% for each category, showing the maximum difference of 1.71%. In the case of U-net, the values for each category were 89.77%, 92.27%, 92.59%, and 92.66%, with the maximum difference of 2.90%. Such results indicate that neither model showed significant differences in detection performance by the characteristics of oil spill distribution. However, the difference in detection tendency was caused by the difference in the model structure and the oil spill distribution characteristics. In all four oil spill categories, the Simple CNN showed a tendency to overestimate the oil spill area and the U-net showed a tendency to underestimate it. These tendencies were emphasized when the border between oil and seawater was unclear.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.20 no.9
• /
• pp.1649-1656
• /
• 2016

Performance of the existing sensor-based fire detection system is limited according to factors in the environment surrounding the sensor. A number of image-based fire detection systems were introduced in order to solve these problem. But such a system can generate a false alarm for objects similar in appearance to fire due to algorithm that directly defines the characteristics of a flame. Also fir detection systems using movement between video flames cannot operate correctly as intended in an environment in which the network is unstable. In this paper, we propose an image-based fire detection method using CNN (Convolutional Neural Network). In this method, firstly we extract fire candidate region using color information from video frame input and then detect fire using trained CNN. Also, we show that the performance is significantly improved compared to the detection rate and missing rate found in previous studies.

• Proceedings of the Korean Society for Agricultural Machinery Conference
• /
• 2017.04a
• /
• pp.173-173
• /
• 2017

드론의 시장규모가 커짐에 따라 초창기 군사 목적에서 현재 민간부문으로 확대되고 있다. 현재 드론은 실외에서 사용될 목적으로 제작된 것이 많으나 실내에서도 드론의 활용 여부가 증가할 것으로 예상된다. 본 연구에서는 실외에서만 사용 가능한 GPS를 대신하여 영상 촬영으로 획득한 이미지를 CNN으로 학습을 시켜 자율고도제어비행을 하도록 한다. 첫 번째로 수동 조작하는 드론에 IMU센서를 부착하여 획득한 고도 데이터를 표로 제시함으로써 GPS를 사용하지 않는 드론의 실내주행에서 일정한 고도 유지는 다소 무리가 있음을 보여준다. 두 번째로 드론의 수동 조작은 일정하지 않은 고도 때문에 CNN의 학습할 영상 획득이 어렵다. 일정한 고도의 영상 획득을 위한 실험용 높이 조절 Base를 제작하여 고도별 영상을 획득한다. 획득한 영상을 통해 얻은 이미지를 CNN 학습을 시킨 후, 학습에 사용되지 않은 이미지를 사용하여 고도 판별을 확인한다. 대조군으로 실내장소를 바꾸어 미리 학습된 CNN으로 고도 판별을 확인한다. 학습에 사용된 이미지의 환경(생명공학관)과 대조군(제 2 공학관)이 촬영된 장소의 환경요소의 차이로 오차가 발생한다. 오차는 실내 장소의 총 높이의 차이 및 서로 상이한 천장 구조물에 따른 것으로 사료되며 Data crop을 통해 획득한 이미지의 천정 부분을 제거하여 노이즈를 줄여 고도 판별의 정확도를 높일 수 있을 것으로 예상한다. 세 번째, CNN으로 학습을 통해 Model을 도출하여 자율 고도 제어 프로세스를 제시한다. 그리고 해당 프로세스를 이용한 자율고도제어 주행과 수동조작을 통한 주행에서의 Z축 가속도 데이터의 표준편차를 비교하여 본 연구의 실효성을 보여준다

• Convergence Security Journal
• /
• v.20 no.3
• /
• pp.53-60
• /
• 2020

At the present stage of the fourth industrial revolution, machine learning and artificial intelligence technologies are rapidly developing, and there is a movement to apply machine learning technology in the security field. Malicious code, including new and transformed, generates an average of 390,000 a day worldwide. Statistics show that security companies ignore or miss 31 percent of alarms. As many malicious codes are generated, it is becoming difficult for humans to detect all malicious codes. As a result, research on the detection of malware and network intrusion events through machine learning is being actively conducted in academia and industry. In international conferences and journals, research on security data analysis using deep learning, a field of machine learning, is presented. have. However, these papers focus on detection accuracy and modify several parameters to improve detection accuracy but do not consider the ratio of dataset. Therefore, this paper aims to reduce the cost and resources of many machine learning research by finding the ratio of dataset that can derive the highest detection accuracy in CNN Mobile net-based malware detection model.

• Journal of Korea Multimedia Society
• /
• v.22 no.10
• /
• pp.1168-1177
• /
• 2019

In order to diagnose and prevent Alzheimer's Disease (AD), it is becoming increasingly important to develop a CAD(Computer-aided Diagnosis) system for AD diagnosis, which provides effective treatment for patients by analyzing 3D MRI images. It is essential to apply powerful deep learning algorithms in order to automatically classify stages of Alzheimer's Disease and to develop a Alzheimer's Disease support diagnosis system that has the function of detecting hippocampus and CSF(Cerebrospinal fluid) which are important biomarkers in diagnosis of Alzheimer's Disease. In this paper, for AD diagnosis, we classify a given MRI data into three categories of AD, mild cognitive impairment, and normal control according by applying 3D brain MRI image to the Faster R-CNN model and detect hippocampus and CSF in MRI image. To do this, we use the 2D MRI slice images extracted from the 3D MRI data of the Faster R-CNN, and perform the widely used majority voting algorithm on the resulting bounding box labels for classification. To verify the proposed method, we used the public ADNI data set, which is the standard brain MRI database. Experimental results show that the proposed method achieves impressive classification performance compared with other state-of-the-art methods.

• Journal of IKEEE
• /
• v.22 no.4
• /
• pp.1079-1087
• /
• 2018

Recently, computer vision field based deep learning artificial intelligence has become a hot topic among various image analysis boundaries. In this study, flames are detected in fire images using the Faster R-CNN algorithm, which is used to detect objects within the image, among various image recognition algorithms based on deep learning. In order to improve fire detection accuracy through a small amount of data sets in the learning process, we use image augmentation techniques, and learn image augmentation by dividing into 6 types and compare accuracy, precision and detection rate. As a result, the detection rate increases as the type of image augmentation increases. However, as with the general accuracy and detection rate of other object detection models, the false detection rate is also increased from 10% to 30%.