• Nutrition Research and Practice
• /
• v.13 no.6
• /
• pp.521-528
• /
• 2019

BACKGROUND/OBJECTIVES: The aim of this study was to develop Korean food image detection and recognition model for use in mobile devices for accurate estimation of dietary intake. MATERIALS/METHODS: We collected food images by taking pictures or by searching web images and built an image dataset for use in training a complex recognition model for Korean food. Augmentation techniques were performed in order to increase the dataset size. The dataset for training contained more than 92,000 images categorized into 23 groups of Korean food. All images were down-sampled to a fixed resolution of $150{\times}150$ and then randomly divided into training and testing groups at a ratio of 3:1, resulting in 69,000 training images and 23,000 test images. We used a Deep Convolutional Neural Network (DCNN) for the complex recognition model and compared the results with those of other networks: AlexNet, GoogLeNet, Very Deep Convolutional Neural Network, VGG and ResNet, for large-scale image recognition. RESULTS: Our complex food recognition model, K-foodNet, had higher test accuracy (91.3%) and faster recognition time (0.4 ms) than those of the other networks. CONCLUSION: The results showed that K-foodNet achieved better performance in detecting and recognizing Korean food compared to other state-of-the-art models.

• Smart Structures and Systems
• /
• v.29 no.1
• /
• pp.195-206
• /
• 2022

Crack detection is essential for inspection of existing structures and crack segmentation based on deep learning is a significant solution. However, datasets are usually one of the key issues. When building a new dataset for deep learning, laborious and time-consuming annotation of a large number of crack images is an obstacle. The aim of this study is to develop an approach that can automatically select a small portion of the most informative crack images from a large pool in order to annotate them, not to label all crack images. An active learning method with difficulty learning mechanism for crack segmentation tasks is proposed. Experiments are carried out on a crack image dataset of a steel box girder, which contains 500 images of 320×320 size for training, 100 for validation, and 190 for testing. In active learning experiments, the 500 images for training are acted as unlabeled image. The acquisition function in our method is compared with traditional acquisition functions, i.e., Query-By-Committee (QBC), Entropy, and Core-set. Further, comparisons are made on four common segmentation networks: U-Net, DeepLabV3, Feature Pyramid Network (FPN), and PSPNet. The results show that when training occurs with 200 (40%) of the most informative crack images that are selected by our method, the four segmentation networks can achieve 92%-95% of the obtained performance when training takes place with 500 (100%) crack images. The acquisition function in our method shows more accurate measurements of informativeness for unlabeled crack images compared to the four traditional acquisition functions at most active learning stages. Our method can select the most informative images for annotation from many unlabeled crack images automatically and accurately. Additionally, the dataset built after selecting 40% of all crack images can support crack segmentation networks that perform more than 92% when all the images are used.

• Journal of the Semiconductor & Display Technology
• /
• v.22 no.1
• /
• pp.11-17
• /
• 2023

This paper proposes a lightweight deep learning network for detecting an image splicing forgery. The research on image forgery detection using CNN, a deep learning network, and research on detecting and localizing forgery in pixel units are in progress. Among them, CAT-Net, which learns the discrete cosine transform coefficients of images together with images, was released in 2022. The DCT coefficients presented by CAT-Net are combined with the JPEG artifact learning module and the backbone model as pre-learning, and the weights are fixed. The dataset used for pre-training is not included in the public dataset, and the backbone model has a relatively large number of network parameters, which causes overfitting in a small dataset, hindering generalization performance. In this paper, this learning module is designed to learn the characterization depending on the DCT domain in real-time during network training without pre-training. The DCT RRU-Net proposed in this paper is a network that combines RRU-Net which detects forgery by learning only images and JPEG artifact learning module. It is confirmed that the network parameters are less than those of CAT-Net, the detection performance of forgery is better than that of RRU-Net, and the generalization performance for various datasets improves through the network architecture and training method of DCT RRU-Net.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.14 no.12
• /
• pp.4664-4681
• /
• 2020

Modulation recognition (MR) plays a key role in cognitive radar, cognitive radio, and some other civilian and military fields. While existing methods can identify the signal modulation type by extracting the signal characteristics, the quality of feature extraction has a serious impact on the recognition results. In this paper, an end-to-end MR method based on long short-term memory (LSTM) and the gated recurrent unit (GRU) is put forward, which can directly predict the modulation type from a sampled signal. Additionally, the sliding window method is applied to fast-changing mixed-modulation signals for which the signal modulation type changes over time. The recognition accuracy on training datasets in different SNR ranges and the proportion of each modulation method in misclassified samples are analyzed, and it is found to be reasonable to select the evenly-distributed and full range of SNR data as the training data. With the improvement of the SNR, the recognition accuracy increases rapidly. When the length of the training dataset increases, the neural network recognition effect is better. The loss function value of the neural network decreases with the increase of the training dataset length, and then tends to be stable. Moreover, when the fast-changing period is less than 20ms, the error rate is as high as 50%. As the fast-changing period is increased to 30ms, the error rates of the GRU and LSTM neural networks are less than 5%.

• Journal of Korea Multimedia Society
• /
• v.25 no.2
• /
• pp.237-246
• /
• 2022

Among the technologies of the 4th industrial revolution, drones that have grown rapidly and are being used in various industries can be operated by the pilot directly or can be operated automatically through programming. In order to be controlled by a pilot or to operate automatically, it is essential to predict and analyze the optimal path for the drone to move without obstacles. In this paper, after securing and analyzing the pilot training dataset through the unmanned aerial vehicle piloting training platform designed through prior research, the profile of the dataset that should be preceded to search and derive the optimal route of the unmanned aerial vehicle was designed. The drone pilot training data includes the speed, movement distance, and angle of the drone, and the data set is visualized to unify the properties showing the same pattern into one and preprocess the properties showing the outliers. It is expected that the proposed big data-based profile can be used to predict and analyze the optimal movement path of an unmanned aerial vehicle.

• The Transactions of The Korean Institute of Electrical Engineers
• /
• v.60 no.2
• /
• pp.398-406
• /
• 2011

In this paper, we introduce a new architecture of PSO-based Polynomial Neural Networks (PNN) and discuss its comprehensive design methodology. The conventional PNN is based on a extended Group Method of Data Handling (GMDH) method, and utilized the polynomial order (viz. linear, quadratic, and modified quadratic) as well as the number of node inputs fixed (selected in advance by designer) at Polynomial Neurons located in each layer through a growth process of the network. Moreover it does not guarantee that the conventional PNN generated through learning results in the optimal network architecture. The PSO-based PNN results in a structurally optimized structure and comes with a higher level of flexibility that the one encountered in the conventional PNN. The PSO-based design procedure being applied at each layer of PNN leads to the selection of preferred PNs with specific local characteristics (such as the number of input variables, input variables, and the order of the polynomial) available within the PNN. In the sequel, two general optimization mechanisms of the PSO-based PNN are explored: the structural optimization is realized via PSO whereas in case of the parametric optimization we proceed with a standard least square method-based learning. To evaluate the performance of the PSO-based PNN, the model is experimented with using Gas furnace process data, and pH neutralization process data. For the characteristic analysis of the given entire data with non-linearity and the construction of efficient model, the given entire system data is partitioned into two type such as Division I(Training dataset and Testing dataset) and Division II(Training dataset, Validation dataset, and Testing dataset). A comparative analysis shows that the proposed PSO-based PNN is model with higher accuracy as well as more superb predictive capability than other intelligent models presented previously.

• Journal of Korean Association for Spatial Structures
• /
• v.21 no.4
• /
• pp.39-48
• /
• 2021

In this paper, a GAN-based data augmentation method is proposed for topology optimization. In machine learning techniques, a total amount of dataset determines the accuracy and robustness of the trained neural network architectures, especially, supervised learning networks. Because the insufficient data tends to lead to overfitting or underfitting of the architectures, a data augmentation method is need to increase the amount of data for reducing overfitting when training a machine learning model. In this study, the Ganerative Adversarial Network (GAN) is used to augment the topology optimization dataset. The produced dataset has been compared with the original dataset.

• The Transactions of the Korea Information Processing Society
• /
• v.13 no.4
• /
• pp.181-188
• /
• 2024

Traditional profanity detection methods have limitations in identifying intentionally altered profanities. This paper introduces a new method based on Named Entity Recognition, a subfield of Natural Language Processing. We developed a profanity detection technique using sequence labeling, for which we constructed a dataset by labeling some profanities in Korean malicious comments and conducted experiments. Additionally, to enhance the model's performance, we augmented the dataset by labeling parts of a Korean hate speech dataset using one of the large language models, ChatGPT, and conducted training. During this process, we confirmed that filtering the dataset created by the large language model by humans alone could improve performance. This suggests that human oversight is still necessary in the dataset augmentation process.

• Journal of Biomedical Engineering Research
• /
• v.43 no.1
• /
• pp.19-26
• /
• 2022

The Comprehensive in vitro Proarrhythmic Assay(CiPA) project was launched for solving the hERG assay problem of being classified as high-risk groups even though they are low-risk drugs due to their high sensitivity. CiPA presented a protocol to predict drug toxicity using physiological data calculated based on the in-silico model. in this study, features calculated through the in-silico model are analyzed for correlation of changing action potential in the near future, and features are verified through predictive performance according to drug datasets. Using the O'Hara Rudy model modified by Dutta et al., Pearson correlation analysis was performed between 13 features(dVm/dtmax, APpeak, APresting, APD90, APD50, APDtri, Ca_peak, Ca_resting, CaD90, CaD50, CaDtri, qNet, qInward) calculated at 100 pacing, and between dVm/d_{tmax_repol} calculated at 1,000 pacing, and linear regression analysis was performed on each of the 12 training drugs, 16 verification drugs, and 28 drugs. Indicators showing high coefficient of determination(R²) in the training drug dataset were qNet 0.93, AP resting 0.83, APDtri 0.78, Ca resting 0.76, dVm/dt_max 0.63, and APD90 0.61. The indicators showing high determinants in the validated drug dataset were APDtri 0.94, APD90 0.92, APD50 0.85, CaD50 0.84, qNet 0.76, and CaD90 0.64. Indicators with high coefficients of determination for all 28 drugs are qNet 0.78, APD90 0.74, and qInward 0.59. The indicators vary in predictive performance depending on the drug dataset, and qNet showed the same high performance of 0.7 or more on the training drug dataset, the verified drug dataset, and the entire drug dataset.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.14 no.7
• /
• pp.2752-2768
• /
• 2020

Phishing websites can have devastating effects on governmental, financial, and social services, as well as on individual privacy. Currently, many phishing detection solutions are evaluated using small datasets and, thus, are prone to sampling issues, such as representing legitimate websites by only high-ranking websites, which could make their evaluation less relevant in practice. Phishing detection solutions which depend only on the URL are attractive, as they can be used in limited systems, such as with firewalls. In this paper, we present a URL-only phishing detection solution based on a convolutional neural network (CNN) model. The proposed CNN takes the URL as the input, rather than using predetermined features such as URL length. For training and evaluation, we have collected over two million URLs in a massive URL phishing detection (MUPD) dataset. We split MUPD into training, validation and testing datasets. The proposed CNN achieves approximately 96% accuracy on the testing dataset; this accuracy is achieved with URL schemes (such as HTTP and HTTPS) removed from the URL. Our proposed solution achieved better accuracy compared to an existing state-of-the-art URL-only model on a published dataset. Finally, the results of our experiment suggest keeping the CNN up-to-date for better results in practice.