• Journal of the Semiconductor & Display Technology
• /
• v.20 no.2
• /
• pp.126-130
• /
• 2021

The ball grid array is one of the packaging methods that used in high density printed circuit board. Solder void defects caused by voids in the solder ball during the BGA process do not directly affect the reliability of the product, but it may accelerate the aging of the device on the PCB layer or interface surface depending on its size or location. Void inspection is important because it is related in yields with products. The most important process in the optical inspection of solder void is the segmentation process of solder and void. However, there are several segmentation algorithms for the vision inspection, it is impossible to inspect all of images ideally. When X-Ray images with poor contrast and high level of noise become difficult to perform image processing for vision inspection in terms of software programming. This paper suggests the solution to deal with the suggested problem by means of using Mask R-CNN instead of digital image processing algorithm. Mask R-CNN model can be trained with images pre-processed to increase contrast or alleviate noises. With this process, it provides more efficient system about complex object segmentation than conventional system.

• Smart Structures and Systems
• /
• v.28 no.6
• /
• pp.827-838
• /
• 2021

The dynamic displacement is considered to be an important indicator of structural safety, and becomes an indispensable part of Structural Health Monitoring (SHM) system for high-speed railway bridges. This paper proposes an indirect strain based dynamic displacement reconstruction methodology for high-speed railway box girders. For the typical box girders under eccentric train load, the plane section assumption and elementary beam theory is no longer applicable due to the bend-torsion coupling effects. The monitored strain was decoupled into bend and torsion induced strain, pre-trained multi-output support vector regression (M-SVR) model was employed for such decoupling process considering the sensor layout cost and reconstruction accuracy. The decoupled strained based displacement could be reconstructed respectively using box girder plate element analysis and mode superposition principle. For the transformation modal matrix has a significant impact on the reconstructed displacement accuracy, the modal order would be optimized using particle swarm algorithm (PSO), aiming to minimize the ill conditioned degree of transformation modal matrix and the displacement reconstruction error. Numerical simulation and dynamic load testing results show that the reconstructed displacement was in good agreement with the simulated or measured results, which verifies the validity and accuracy of the algorithm proposed in this paper.

• Journal of Korea Multimedia Society
• /
• v.24 no.12
• /
• pp.1598-1605
• /
• 2021

Recently, convolutional neural networks (CNNs) have been widely used with excellent performance in various computer vision fields, including super-resolution (SR). However, CNN is computationally intensive and requires a lot of memory, making it difficult to apply to limited hardware resources such as mobile or Internet of Things devices. To solve these limitations, network lightening studies have been actively conducted to reduce the depth or size of pre-trained deep CNN models while maintaining their performance as much as possible. This paper aims to lighten the SR CNN model, SRGAN, using the knowledge distillation among network lightening technologies; thus, it proposes four techniques with different methods of transferring the knowledge of the teacher network to the student network and presents experiments to compare and analyze the performance of each technique. In our experimental results, it was confirmed through quantitative and qualitative evaluation indicators that student networks with knowledge transfer performed better than those without knowledge transfer, and among the four knowledge transfer techniques, the technique of conducting adversarial learning after transferring knowledge from the teacher generator to the student generator showed the best performance.

• International Journal of Computer Science & Network Security
• /
• v.22 no.9
• /
• pp.23-30
• /
• 2022

Glaucoma is a chronic neuropathy that affects the optic nerve which can lead to blindness. The detection and prediction of glaucoma become possible using deep neural networks. However, the detection performance relies on the availability of a large number of data. Therefore, we propose different frameworks, including a hybrid of a generative adversarial network and a convolutional neural network to automate and increase the performance of glaucoma detection. The proposed frameworks are evaluated using five public glaucoma datasets. The framework which uses a Deconvolutional Generative Adversarial Network (DCGAN) and a DenseNet pre-trained model achieves 99.6%, 99.08%, 99.4%, 98.69%, and 92.95% of classification accuracy on RIMONE, Drishti-GS, ACRIMA, ORIGA-light, and HRF datasets respectively. Based on the experimental results and evaluation, the proposed framework closely competes with the state-of-the-art methods using the five public glaucoma datasets without requiring any manually preprocessing step.

• Journal of Information Processing Systems
• /
• v.19 no.1
• /
• pp.109-117
• /
• 2023

Defect detection is one of the key factors in fabric quality control. To improve the speed and accuracy of denim fabric defect detection, this paper proposes a defect detection algorithm based on cascading feature extraction architecture. Firstly, this paper extracts these weight parameters of the pre-trained VGG16 model on the large dataset ImageNet and uses its portability to train the defect detection classifier and the defect recognition classifier respectively. Secondly, retraining and adjusting partial weight parameters of the convolution layer were retrained and adjusted from of these two training models on the high-definition fabric defect dataset. The last step is merging these two models to get the defect detection algorithm based on cascading architecture. Then there are two comparative experiments between this improved defect detection algorithm and other feature extraction methods, such as VGG16, ResNet-50, and Xception. The results of experiments show that the defect detection accuracy of this defect detection algorithm can reach 94.3% and the speed is also increased by 1-3 percentage points.

• Annual Conference on Human and Language Technology
• /
• 2022.10a
• /
• pp.238-244
• /
• 2022

감정 분석은 텍스트에 표현된 인간의 감정을 인식하여 다양한 감정 유형으로 분류하는 것이다. 섬세한 인간의 감정을 보다 정확히 분류하기 위해서는 감정 유형의 분류가 무엇보다 중요하다. 본 연구에서는 사전 학습 언어 모델을 활용하여 우리말샘의 감정 어휘와 용례를 바탕으로 기쁨, 슬픔, 공포, 분노, 혐오, 놀람, 흥미, 지루함, 통증의 감정 유형으로 분류된 감정 말뭉치를 구축하였다. 감정 말뭉치를 구축한 후 성능 평가를 위해 대표적인 트랜스포머 기반 사전 학습 모델 중 RoBERTa, MultiDistilBert, MultiBert, KcBert, KcELECTRA. KoELECTRA를 활용하여 보다 넓은 범위에서 객관적으로 모델 간의 성능을 평가하고 각 감정 유형별 정확도를 바탕으로 감정 유형의 특성을 알아보았다. 그 결과 각 모델의 학습 구조가 다중 분류 말뭉치에 어떤 영향을 주는지 구체적으로 파악할 수 있었으며, ELECTRA가 상대적으로 우수한 성능을 보여주고 있음을 확인하였다. 또한 감정 유형별 성능을 비교를 통해 다양한 감정 유형 중 기쁨, 슬픔, 공포에 대한 성능이 우수하다는 것을 알 수 있었다.

• Annual Conference on Human and Language Technology
• /
• 2022.10a
• /
• pp.613-618
• /
• 2022

효과적인 자연어 처리를 위해 제안된 Transformer 구조의 등장 이후, 이를 활용한 대규모 언어 모델이자 사전학습 모델인 BERT, GPT, OPT 등이 공개되었고, 이들을 한국어에 보다 특화한 KoBERT, KoGPT 등의 사전학습 모델이 공개되었다. 자연어 처리 모델의 확보를 위한 학습 자원이 늘어나고 있지만, 사전학습 모델을 각종 응용작업에 적용하기 위해서는 데이터 준비, 코드 작성, 파인 튜닝 및 저장과 같은 복잡한 절차를 수행해야 하며, 이는 다수의 응용 사용자에게 여전히 도전적인 과정으로, 올바른 결과를 도출하는 것은 쉽지 않다. 이러한 어려움을 완화시키고, 다양한 기계 학습 모델을 사용자 데이터에 보다 쉽게 적용할 수 있도록 AutoML으로 통칭되는 자동 하이퍼파라미터 탐색, 모델 구조 탐색 등의 기법이 고안되고 있다. 본 연구에서는 한국어 사전학습 모델과 한국어 텍스트 데이터를 사용한 자연어 처리 모델 산출 과정을 정형화 및 절차화하여, 궁극적으로 목표로 하는 예측 모델을 자동으로 산출하는 시스템의 설계를 소개한다.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.18 no.4
• /
• pp.693-700
• /
• 2023

Sentiment analysis, a branch of natural language processing that classifies and identifies subjective opinions and emotions in text documents as positive or negative, can be used for various promotions and services through customer preference analysis. To this end, recent research has been conducted utilizing various techniques in machine learning and deep learning. In this study, we propose an optimal language model by comparing the accuracy of sentiment analysis for movie, product, and game reviews using existing RNN-based models and recent Transformer-based language models. In our experiments, LMKorBERT and GPT3 showed relatively good accuracy among the models pre-trained on the Korean corpus.

• International journal of advanced smart convergence
• /
• v.12 no.2
• /
• pp.25-33
• /
• 2023

The purpose of this study is to evaluate the impact of intermediate features on FER performance. To achieve this objective, intermediate features were extracted from the input images at specific layers (FM1~FM4) of the pre-trained network (Resnet-18). These extracted intermediate features and original images were used as inputs to the vision transformer (ViT), and the FER performance was compared. As a result, when using a single image as input, using intermediate features extracted from FM2 yielded the best performance (training accuracy: 94.35%, testing accuracy: 75.51%). When using the original image as input, the training accuracy was 91.32% and the testing accuracy was 74.68%. However, when combining the original image with intermediate features as input, the best FER performance was achieved by combining the original image with FM2, FM3, and FM4 (training accuracy: 97.88%, testing accuracy: 79.21%). These results imply that incorporating intermediate features alongside the original image can lead to superior performance. The findings can be referenced and utilized when designing the preprocessing stages of a deep learning model in FER. By considering the effectiveness of using intermediate features, practitioners can make informed decisions to enhance the performance of FER systems.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2020.11a
• /
• pp.971-974
• /
• 2020

Infants express their physical and emotional needs to the outside world mainly through crying. However, most of parents find it challenging to understand the reason behind their babies' cries. Failure to correctly understand the cause of a baby' cry and take appropriate actions can affect the cognitive and motor development of newborns undergoing rapid brain development. In this paper, we propose an infant cry recognition system based on deep transfer learning to help parents identify crying babies' needs the same way a specialist would. The proposed system works by transforming the waveform of the cry signal into log-mel spectrogram, then uses the VGGish model pre-trained on AudioSet to extract a 128-dimensional feature vector from the spectrogram. Finally, a softmax function is used to classify the extracted feature vector and recognize the corresponding type of cry. The experimental results show that our method achieves a good performance exceeding 0.96 in precision and recall, and f1-score.