• Journal of Platform Technology
• /
• v.8 no.4
• /
• pp.20-28
• /
• 2020

In this paper, a voice, object, and optical character recognition platform including voice recognition-based smart wearable devices, smart devices, and web AI servers was proposed as an appropriate technology to help the visually impaired to live independently by learning the life cycle of the visually impaired in advance. The wearable device for the visually impaired was designed and manufactured with a reverse neckband structure to increase the convenience of wearing and the efficiency of object recognition. And the high-sensitivity small microphone and speaker attached to the wearable device was configured to support the voice recognition interface function consisting of the app of the smart device linked to the wearable device. From experimental results, the voice, object, and optical character recognition service used open source and Google APIs in the web AI server, and it was confirmed that the accuracy of voice, object and optical character recognition of the service platform achieved an average of 90% or more.

• The Journal of Korean Institute of Information Technology
• /
• v.17 no.5
• /
• pp.13-20
• /
• 2019

Recently, object detection technology based on CNN has been actively studied. Object detection technology is used as an important technology in autonomous vehicles, intelligent image analysis, and so on. In this paper, we propose a rotation change robust object detection system by applying MI-FL (Moment Invariant-Feature Layer) to SSD (Single Shot Multibox Detector) which is one of CNN-based object detectors. First, the features of the input image are extracted based on the VGG network. Then, a total of six feature layers are applied to generate bounding boxes by predicting the location and type of object. We then use the NMS algorithm to get the bounding box that is the most likely object. Once an object bounding box has been determined, the invariant moment feature of the corresponding region is extracted using MI-FL, and stored and learned in advance. In the detection process, it is possible to detect the rotated image more robust than the conventional method by using the previously stored moment invariant feature information. The performance improvement of about 4 ~ 5% was confirmed by comparing SSD with existing SSD and MI-FL.

• Journal of the Korea Convergence Society
• /
• v.12 no.1
• /
• pp.29-35
• /
• 2021

License plate recognition plays a key role in intelligent transportation systems. Therefore, it is a very important process to efficiently detect the number and character areas. In this paper, we propose a method to effectively detect license plate number area by applying deep learning and semantic image segmentation algorithm. The proposed method is an algorithm that detects number and text areas directly from the license plate without preprocessing such as pixel projection. The license plate image was acquired from a fixed camera installed on the road, and was used in various real situations taking into account both weather and lighting changes. The input images was normalized to reduce the color change, and the deep learning neural networks used in the experiment were Vgg16, Vgg19, ResNet18, and ResNet50. To examine the performance of the proposed method, we experimented with 500 license plate images. 300 sheets were used for learning and 200 sheets were used for testing. As a result of computer simulation, it was the best when using ResNet50, and 95.77% accuracy was obtained.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.20 no.6
• /
• pp.51-58
• /
• 2020

The blind people have difficulty living an independent clothing life. The furniture and home appliance are adding AI or IoT with the recent growth of the smart appliance market. To support the independent clothing life of the blind, this paper suggests a smart wardrobe with closet control function, voice recognition function and clothes information recognition using CNN algorithm. The number of layers of the model was changed and Maxpooling was adjusted to create the model to increase accuracy in the process of recognizing clothes. Early Stopping Callback option is applied to ensure learning accuracy when creating a model. We added Dropout to prevent overfitting. The final model created by this process can be found to have 80 percent accuracy in clothing recognition.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.30 no.6
• /
• pp.1031-1041
• /
• 2020

Cloud computing technology plays an important role in the deep learning industry as deep learning services are deployed frequently on top of cloud infrastructures. In such cloud environment, virtualization technology provides logically independent and isolated computing space for each tenant. However, recent studies demonstrate that by leveraging vulnerabilities of virtualization techniques and shared processor architectures in the cloud system, various side-channels can be established between cloud tenants. In this paper, we propose a novel attack scenario that can steal internal information of deep learning models by exploiting the Meltdown vulnerability in a multi-tenant system environment. On the basis of our experiment, the proposed attack method could extract internal information of a TensorFlow deep-learning service with 92.875% accuracy and 1.325kB/s extraction speed.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.24 no.9
• /
• pp.1224-1230
• /
• 2020

Since the spread of smart phones, interest in wearable devices has increased and diversified, and is closely related to the lives of users, and has been used as a method for providing personalized services. In this paper, we propose a method to detect the user's behavior by applying information from a 3-axis acceleration sensor and a 3-axis gyro sensor embedded in a smartphone to a convolutional neural network. Human behavior differs according to the size and range of motion, starting and ending time, including the duration of the signal data constituting the motion. Therefore, there is a performance problem for accuracy when applied to a convolutional neural network as it is. Therefore, we proposed a Time-Division Feature Fusion Convolutional Neural Network (TDFFCNN) that learns the characteristics of the sensor data segmented over time. The proposed method outperformed other classifiers such as SVM, IBk, convolutional neural network, and long-term memory circulatory neural network.

• Journal of Convergence for Information Technology
• /
• v.11 no.3
• /
• pp.7-13
• /
• 2021

This study develops an artificial intelligence prediction system for Fine particulate Matter(PM2.5) based on the deep learning algorithm GAN model. The experimental data are closely related to the changes in temperature, humidity, wind speed, and atmospheric pressure generated by the time series axis and the concentration of air pollutants such as SO2, CO, O3, NO2, and PM10. Due to the characteristics of the data, since the concentration at the current time is affected by the concentration at the previous time, a predictive model for recursive supervised learning was applied. For comparative analysis of the accuracy of the existing models, CNN and LSTM, the difference between observation value and prediction value was analyzed and visualized. As a result of performance analysis, it was confirmed that the proposed GAN improved to 15.8%, 10.9%, and 5.5% in the evaluation items RMSE, MAPE, and IOA compared to LSTM, respectively.

• The Journal of the Korea Contents Association
• /
• v.21 no.8
• /
• pp.1-9
• /
• 2021

Deep learning in computer vision has made accelerated improvement over a short period but large-scale learning data and computing power are still essential that required time-consuming trial and error tasks are involved to derive an optimal network model. In this study, we propose a similar image classification performance improvement method based on CR (Confusion Rate) that considers only the characteristics of the data itself regardless of network optimization or data reinforcement. The proposed method is a technique that improves the performance of the deep learning model by calculating the CRs for images in a dataset with similar characteristics and reflecting it in the weight of the Loss Function. Also, the CR-based recognition method is advantageous for image identification with high similarity because it enables image recognition in consideration of similarity between classes. As a result of applying the proposed method to the Resnet18 model, it showed a performance improvement of 0.22% in HanDB and 3.38% in Animal-10N. The proposed method is expected to be the basis for artificial intelligence research using noisy labeled data accompanying large-scale learning data.

• Design & Manufacturing
• /
• v.15 no.4
• /
• pp.24-31
• /
• 2021

In this study, an artificial neural network model was constructed to convert CAE analysis data into similar experimental data. In the analysis and experiment, the injection molding data for 50 conditions were acquired through the design of experiment and random selection method. The injection molding conditions and the weight, height, and diameter of the product derived from CAE results were used as the input parameters for learning of the convert model. Also the product qualities of experimental results were used as the output parameters for learning of the convert model. The accuracy of the convert model showed RMSE values of 0.06g, 0.03mm, and 0.03mm in weight, height, and diameter, respectively. As the next step, additional randomly selected conditions were created and CAE analysis was performed. Then, the additional CAE analysis data were converted to similar experimental data through the conversion model. An artificial neural network model was constructed to predict the quality of injection molded product by using converted similar experimental data and injection molding experiment data. The injection molding conditions were used as input parameters for learning of the predicted model and weight, height, and diameter of the product were used as output parameters for learning. As a result of evaluating the performance of the prediction model, the predicted weight, height, and diameter showed RMSE values of 0.11g, 0.03mm, and 0.05mm and in terms of quality criteria of the target product, all of them showed accurate results satisfying the criteria range.

• The Journal of The Korea Institute of Intelligent Transport Systems
• /
• v.20 no.6
• /
• pp.242-263
• /
• 2021

This paper presents a method to implement a low-cost driving simulator for developing autonomous driving algorithms. This is implemented by using GTA V, a physical engine-based commercial game software, containing a function to emulate output and data of various sensors for autonomous driving. For this, NF of Script Hook V is incorporated to acquire GT data by accessing internal data of the software engine, and then, various sensor data for autonomous driving are generated. We present an overall function of the developed driving simulator and perform a verification of individual functions. We explain the process of acquiring GT data via direct access to the internal memory of the game engine to build up an autonomous driving algorithm development environment. And, finally, an example applicable to artificial neural network training and performance evaluation by processing the emulated sensor output is included.