• Journal of Korea Multimedia Society
• /
• v.25 no.10
• /
• pp.1375-1385
• /
• 2022

In this paper, we propose a semi-supervised domain adaptation solution to deal with practical face recognition (FR) scenarios where a single face image for each target identity (to be recognized) is only available in the training phase. Main goal of the proposed method is to reduce the discrepancy between the target and the source domain face images, which ultimately improves FR performances. The proposed method is based on the Domain Adatation network (DAN) using an MMD loss function to reduce the discrepancy between domains. In order to train more effectively, we develop a novel loss function learning strategy in which MMD loss and cross-entropy loss functions are adopted by using different weights according to the progress of each epoch during the learning. The proposed weight adoptation focuses on the training of the source domain in the initial learning phase to learn facial feature information such as eyes, nose, and mouth. After the initial learning is completed, the resulting feature information is used to training a deep network using the target domain images. To evaluate the effectiveness of the proposed method, FR performances were evaluated with pretrained model trained only with CASIA-webface (source images) and fine-tuned model trained only with FERET's gallery (target images) under the same FR scenarios. The experimental results showed that the proposed semi-supervised domain adaptation can be improved by 24.78% compared to the pre-trained model and 28.42% compared to the fine-tuned model. In addition, the proposed method outperformed other state-of-the-arts domain adaptation approaches by 9.41%.

• Smart Structures and Systems
• /
• v.31 no.4
• /
• pp.335-349
• /
• 2023

Bridges constantly undergo deterioration and damage, the most common ones being concrete damage and exposed rebar. Periodic inspection of bridges to identify damages can aid in their quick remediation. Likewise, identifying components can provide context for damage assessment and help gauge a bridge's state of interaction with its surroundings. Current inspection techniques rely on manual site visits, which can be time-consuming and costly. More recently, robotic inspection assisted by autonomous data analytics based on Computer Vision (CV) and Artificial Intelligence (AI) has been viewed as a suitable alternative to manual inspection because of its efficiency and accuracy. To aid research in this avenue, this study performs a comparative assessment of different architectures, loss functions, and ensembling strategies for the autonomous segmentation of bridge components and damages. The experiments lead to several interesting discoveries. Nested Reg-UNet architecture is found to outperform five other state-of-the-art architectures in both damage and component segmentation tasks. The architecture is built by combining a Nested UNet style dense configuration with a pretrained RegNet encoder. In terms of the mean Intersection over Union (mIoU) metric, the Nested Reg-UNet architecture provides an improvement of 2.86% on the damage segmentation task and 1.66% on the component segmentation task compared to the state-of-the-art UNet architecture. Furthermore, it is demonstrated that incorporating the Lovasz-Softmax loss function to counter class imbalance can boost performance by 3.44% in the component segmentation task over the most employed alternative, weighted Cross Entropy (wCE). Finally, weighted softmax ensembling is found to be quite effective when used synchronously with the Nested Reg-UNet architecture by providing mIoU improvement of 0.74% in the component segmentation task and 1.14% in the damage segmentation task over a single-architecture baseline. Overall, the best mIoU of 92.50% for the component segmentation task and 84.19% for the damage segmentation task validate the feasibility of these techniques for autonomous bridge component and damage segmentation using RGB images.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.17 no.6
• /
• pp.1516-1529
• /
• 2023

In this paper, a motor gearbox fault-detection system based on a hierarchical flow-based model is proposed. The proposed system is used for the anomaly detection of a motion sound-based actuator module. The proposed flow-based model, which is a generative model, learns by directly modeling a data distribution function. As the objective function is the maximum likelihood value of the input data, the training is stable and simple to use for anomaly detection. The operation sound of a car's side-view mirror motor is converted into a Mel-spectrogram image, consisting of a folding signal and an unfolding signal, and used as training data in this experiment. The proposed system is composed of an encoder and a decoder. The data extracted from the layer of the pretrained feature extractor are used as the decoder input data in the encoder. This information is used in the decoder by performing an interlayer cross-scale convolution operation. The experimental results indicate that the context information of various dimensions extracted from the interlayer hierarchical data improves the defect detection accuracy. This paper is notable because it uses acoustic data and a normalizing flow model to detect outliers based on the features of experimental data.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2023.01a
• /
• pp.35-38
• /
• 2023

포트홀은 주행하는 자동차와 접촉이 이뤄지면 차체나 운전자에게 충격을 주고 제어를 잃게 하여 도로 위 안전을 위협할 수 있다. 포트홀의 검출을 위한 국내 동향으로는 진동을 이용한 방식과 신고시스템 이용한 방식과 영상 인식을 기반한 방식이 있다. 이 중 영상 인식 기반 방식은 보급이 쉽고 비용이 저렴하나, 컴퓨터 비전 알고리즘은 영상의 품질에 따라 정확도가 달라지는 문제가 있었다. 이를 보완하기 위해 영상 인식 기반의 딥러닝 모델을 사용한다. 따라서, 본 논문에서는 사전 학습된 딥러닝 모델의 정확도 향상을 위한 Feature Level Ensemble 기법을 제안한다. 제안된 기법은 사전 학습된 CNN 모델 중 Test 데이터의 정확도 기준 Top-3 모델을 선정하여 각 딥러닝 모델의 Feature Map을 Concatenate하고 이를 Fully-Connected(FC) Layer로 입력하여 구현한다. Feature Level Ensemble 기법이 적용된 딥러닝 모델은 평균 대비 3.76%의 정확도 향상을 보였으며, Top-1 모델인 ShuffleNet보다 0.94%의 정확도 향상을 보였다. 결론적으로 본 논문에서 제안된 기법은 사전 학습된 모델들을 이용하여 각 모델의 다양한 특징을 통해 기존 모델 대비 정확도의 향상을 이룰 수 있었다.

• Korean Journal of Cognitive Science
• /
• v.34 no.3
• /
• pp.197-226
• /
• 2023

Generative Text Summarization is one of the Natural Language Processing tasks. It generates a short abbreviated summary while preserving the content of the long text. ROUGE is a widely used lexical-overlap based metric for text summarization models in generative summarization benchmarks. Although it shows very high performance, the studies report that 30% of the generated summary and the text are still inconsistent. This paper proposes a methodology for evaluating the performance of the summary model without using the correct summary. AggreFACT is a human-annotated dataset that classifies the types of errors in neural text summarization models. Among all the test candidates, the two cases, generation summary, and when errors occurred throughout the summary showed the highest correlation results. We observed that the proposed evaluation score showed a high correlation with models finetuned with BART and PEGASUS, which is pretrained with a large-scale Transformer structure.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.33 no.6
• /
• pp.1087-1098
• /
• 2023

The rapid advancement of artificial intelligence (AI) technology has led to its proactive utilization across various fields. However, this widespread adoption of AI-based systems has raised concerns about the increasing threat of attacks on these systems. In particular, deep neural networks, commonly used in deep learning, have been found vulnerable to adversarial attacks that intentionally manipulate input data to induce model errors. In this study, we propose a method to protect image classification models from visually imperceptible One-Pixel attacks, where only a single pixel is altered in an image. The proposed defense technique utilizes an autoencoder model to remove potential threat elements from input images before forwarding them to the classification model. Experimental results, using the CIFAR-10 dataset, demonstrate that the autoencoder-based defense approach significantly improves the robustness of pretrained image classification models against One-Pixel attacks, with an average defense rate enhancement of 81.2%, all without the need for modifications to the existing models.

• The Journal of Korea Robotics Society
• /
• v.19 no.3
• /
• pp.293-310
• /
• 2024

In this paper, we propose an agent model for Language-Driven Zero-Shot Object Navigation (L-ZSON) tasks, which takes in a freeform language description of an unseen target object and navigates to find out the target object in an inexperienced environment. In general, an L-ZSON agent should able to visually ground the target object by understanding the freeform language description of it and recognizing the corresponding visual object in camera images. Moreover, the L-ZSON agent should be also able to build a rich spatial context map over the unknown environment and decide efficient exploration actions based on the map until the target object is present in the field of view. To address these challenging issues, we proposes AML (Agent Model for L-ZSON), a novel L-ZSON agent model to make effective use of AI foundation models such as Large Language Model (LLM) and Vision-Language model (VLM). In order to tackle the visual grounding issue of the target object description, our agent model employs GLEE, a VLM pretrained for locating and identifying arbitrary objects in images and videos in the open world scenario. To meet the exploration policy issue, the proposed agent model leverages the commonsense knowledge of LLM to make sequential navigational decisions. By conducting various quantitative and qualitative experiments with RoboTHOR, the 3D simulation platform and PASTURE, the L-ZSON benchmark dataset, we show the superior performance of the proposed agent model.

• Nuclear Engineering and Technology
• /
• v.56 no.8
• /
• pp.3123-3128
• /
• 2024

The dicentric chromosome assay is a gold standard method to estimate radiation exposure by calculating the ratio of dicentric chromosomes existing in cells. The objective of this study was to propose an automatic dicentric chromosome discrimination method based on deep convolutional neural networks using radiation exposure patient data. From 45 patients with radiation exposure, conventional Giemsa-stained images of 116,258 normal and 2800 dicentric chromosomes were confirmed. ImageNet was used to pre-train VGG19, which was modified and fine-tuned. The proposed modified VGG19 demonstrated dicentric chromosome discrimination performance, with a true positive rate of 0.927, a true negative rate of 0.997, a positive predictive value of 0.882, a negative predictive value of 0.998, and an area under the receiver operating characteristic curve of 0.997.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.22 no.5
• /
• pp.7-14
• /
• 2021

Tomato crops are highly affected by tomato diseases, and if not prevented, a disease can cause severe losses for the agricultural economy. Therefore, there is a need for a system that quickly and accurately diagnoses various tomato diseases. In this paper, we propose a system that classifies nine diseases as well as healthy tomato plants by applying various pretrained deep learning-based CNN models trained on an ImageNet dataset. The tomato leaf image dataset obtained from PlantVillage is provided as input to ResNet, Xception, and DenseNet, which have deep learning-based CNN architectures. The proposed models were constructed by adding a top-level classifier to the basic CNN model, and they were trained by applying a 5-fold cross-validation strategy. All three of the proposed models were trained in two stages: transfer learning (which freezes the layers of the basic CNN model and then trains only the top-level classifiers), and fine-tuned learning (which sets the learning rate to a very small number and trains after unfreezing basic CNN layers). SGD, RMSprop, and Adam were applied as optimization algorithms. The experimental results show that the DenseNet CNN model to which the RMSprop algorithm was applied output the best results, with 98.63% accuracy.

• Smart Media Journal
• /
• v.10 no.2
• /
• pp.76-83
• /
• 2021

With the development and use of smart devices such as smartphones and tablets increases, the mobile application market based on mobile devices is growing rapidly. Mobile application users write reviews to share their experience in using the application, which can identify consumers' various needs and application developers can receive useful feedback on improving the application through reviews written by consumers. However, there is a need to come up with measures to minimize the amount of time and expense that consumers have to pay to manually analyze the large amount of reviews they leave. In this work, we propose to collect delivery application user reviews from Google PlayStore and then use machine learning and deep learning techniques to classify them into four categories like application feature advantages, disadvantages, feature improvement requests and bug report. In the case of the performance of the Hugging Face's pretrained BERT-based Transformer model, the f1 score values for the above four categories were 0.93, 0.51, 0.76, and 0.83, respectively, showing superior performance than LSTM and GRU.