• 대한건축학회논문집:구조계
• /
• 제36권3호
• /
• pp.113-120
• /
• 2020

The purpose of this study is to compare the models of Deep Learning-based Convolution Neural Network(CNN) for concrete crack detection. The comparison models are AlexNet, GoogLeNet, VGG16, VGG19, ResNet-18, ResNet-50, ResNet-101, and SqueezeNet which won ImageNet Large Scale Visual Recognition Challenge(ILSVRC). To train, validate and test these models, we constructed 3000 training data and 12000 validation data with 256×256 pixel resolution consisting of cracked and non-cracked images, and constructed 5 test data with 4160×3120 pixel resolution consisting of concrete images with crack. In order to increase the efficiency of the training, transfer learning was performed by taking the weight from the pre-trained network supported by MATLAB. From the trained network, the validation data is classified into crack image and non-crack image, yielding True Positive (TP), True Negative (TN), False Positive (FP), False Negative (FN), and 6 performance indicators, False Negative Rate (FNR), False Positive Rate (FPR), Error Rate, Recall, Precision, Accuracy were calculated. The test image was scanned twice with a sliding window of 256×256 pixel resolution to classify the cracks, resulting in a crack map. From the comparison of the performance indicators and the crack map, it was concluded that VGG16 and VGG19 were the most suitable for detecting concrete cracks.

• Imaging Science in Dentistry
• /
• 제53권1호
• /
• pp.27-34
• /
• 2023

Purpose: The aim of this study was to clarify the influence of training with a different kind of lesion on the performance of a target model. Materials and Methods: A total of 310 patients(211 men, 99 women; average age, 47.9±16.1 years) were selected and their panoramic images were used in this study. We created a source model using panoramic radiographs including mandibular radiolucent cyst-like lesions (radicular cyst, dentigerous cyst, odontogenic keratocyst, and ameloblastoma). The model was simulatively transferred and trained on images of Stafne's bone cavity. A learning model was created using a customized DetectNet built in the Digits version 5.0 (NVIDIA, Santa Clara, CA). Two machines(Machines A and B) with identical specifications were used to simulate transfer learning. A source model was created from the data consisting of ameloblastoma, odontogenic keratocyst, dentigerous cyst, and radicular cyst in Machine A. Thereafter, it was transferred to Machine B and trained on additional data of Stafne's bone cavity to create target models. To investigate the effect of the number of cases, we created several target models with different numbers of Stafne's bone cavity cases. Results: When the Stafne's bone cavity data were added to the training, both the detection and classification performances for this pathology improved. Even for lesions other than Stafne's bone cavity, the detection sensitivities tended to increase with the increase in the number of Stafne's bone cavities. Conclusion: This study showed that using different lesions for transfer learning improves the performance of the model.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제15권5호
• /
• pp.1778-1797
• /
• 2021

Document image binarization is an important pre-processing step in document analysis and archiving. The state-of-the-art models for document image binarization are variants of encoder-decoder architectures, such as FCN (fully convolutional network) and U-Net. Despite their success, they still suffer from three limitations: (1) reduced feature map resolution due to consecutive strided pooling or convolutions, (2) multiple scales of target objects, and (3) reduced localization accuracy due to the built-in invariance of deep convolutional neural networks (DCNNs). To overcome these three challenges, we propose an improved semantic segmentation model, referred to as DP-LinkNet, which adopts the D-LinkNet architecture as its backbone, with the proposed hybrid dilated convolution (HDC) and spatial pyramid pooling (SPP) modules between the encoder and the decoder. Extensive experiments are conducted on recent document image binarization competition (DIBCO) and handwritten document image binarization competition (H-DIBCO) benchmark datasets. Results show that our proposed DP-LinkNet outperforms other state-of-the-art techniques by a large margin. Our implementation and the pre-trained models are available at https://github.com/beargolden/DP-LinkNet.

• 한국컴퓨터정보학회논문지
• /
• 제27권11호
• /
• pp.39-46
• /
• 2022

본 논문에서는 UD Korean Kaist v2.3 코퍼스를 이용하여 범용 품사 태깅, 표제어추출 그리고 의존 구문분석을 동시에 예측할 수 있는 보편적 다중 작업 모델을 제안하였다. 제안 모델은 사전학습 언어모델인 다국어 BERT (Multilingual BERT)와 한국어 BERT (KR-BERT와 KoBERT)을 대상으로 추가학습 (fine-tuning)을 수행하여 BERT 모델의 자가-집중 (self-attention) 기법과 그래프 기반 Biaffine attention 기법을 적용하여 제안 모델의 성능을 비교 분석하였다.

• Journal of Information Processing Systems
• /
• 제18권5호
• /
• pp.650-664
• /
• 2022

Mobile applications can be easily downloaded and installed via markets. However, malware and malicious applications containing unwanted advertisements exist in these application markets. Therefore, smartphone users install applications with reference to the application review to avoid such malicious applications. An application review typically comprises contents for evaluation; however, a false review with a specific purpose can be included. Such false reviews are known as fake reviews, and they can be generated using artificial intelligence (AI)-based text-generating models. Recently, AI-based text-generating models have been developed rapidly and demonstrate high-quality generated texts. Herein, we analyze the features of fake reviews generated from Generative Pre-Training-2 (GPT-2), an AI-based text-generating model and create a model to detect those fake reviews. First, we collect a real human-written application review from Kaggle. Subsequently, we identify features of the fake review using natural language processing and statistical analysis. Next, we generate fake review detection models using five types of machine-learning models trained using identified features. In terms of the performances of the fake review detection models, we achieved average F1-scores of 0.738, 0.723, and 0.730 for the fake review, real review, and overall classifications, respectively.

• 정보과학회 논문지
• /
• 제44권3호
• /
• pp.306-313
• /
• 2017

개체명 인식이란 문서 내에서 인명, 기관명, 지명, 시간, 날짜 등 고유한 의미를 가지는 개체명을 추출하여 그 종류를 결정하는 것을 말한다. 최근 개체명 인식 연구에서 가장 우수한 성능을 보여주고 있는 모델은 Bidirectional LSTM CRFs 모델이다. 이러한 LSTM 기반의 딥 러닝 모델은 입력이 되는 단어 표상에 의존적이다. 따라서 입력이 되는 단어를 잘 표현하기 위하여 단어 표상을 확장하는 방법에 대한 연구가 많이 진행되어지고 있다. 본 논문에서는 한국어 개체명 인식을 위하여 Bidirectional LSTM CRFs모델을 사용하고, 그 입력으로 사용되는 단어 표상을 확장하기 위해 사전 학습된 단어 임베딩 벡터, 품사 임베딩 벡터, 음절 기반에서 확장된 단어 임베딩 벡터, 그리고 개체명 사전 자질 벡터를 사용한다. 최종 단어 표상 확장 결과 사전 학습된 단어 임베딩 벡터만 사용한 것 보다 8.05%p의 성능 향상을 보였다.

• 한국정보처리학회:학술대회논문집
• /
• 한국정보처리학회 2019년도 춘계학술발표대회
• /
• pp.472-474
• /
• 2019

Because of the difference in network structure and loss function, Verification and identification models have their respective advantages and limitations for person reidentification (re-ID). In this work, we propose a multi-task network simultaneously computes the identification loss and verification loss for person reidentification. Given a pair of images as network input, the multi-task network simultaneously outputs the identities of the two images and whether the images belong to the same identity. In experiments, we analyze the major factors affect the accuracy of person reidentification. To address the occlusion problem and improve the generalization ability of reID models, we use the Random Erasing Augmentation (REA) method to preprocess the images. The method can be easily applied to different pre-trained networks, such as ResNet and VGG. The experimental results on the Market1501 datasets show significant and consistent improvements over the state-of-the-art methods.

• Journal of Information Processing Systems
• /
• 제17권3호
• /
• pp.537-555
• /
• 2021

Many researchers have carried out studies related to programming languages since the beginning of computer science. Besides programming with traditional programming languages (i.e., procedural, object-oriented, functional programming language, etc.), a new paradigm of programming is being carried out. It is programming with natural language. By programming with natural language, we expect that it will free our expressiveness in contrast to programming languages which have strong constraints in syntax. This paper surveys the approaches that generate source code automatically from a natural language description. We also categorize the approaches by their forms of input and output. Finally, we analyze the current trend of approaches and suggest the future direction of this research domain to improve automatic code generation with natural language. From the analysis, we state that researchers should work on customizing language models in the domain of source code and explore better representations of source code such as embedding techniques and pre-trained models which have been proved to work well on natural language processing tasks.

• Journal of Information Processing Systems
• /
• 제18권4호
• /
• pp.524-534
• /
• 2022

In this paper, we produce Trash Object Detection (TOD) dataset to solve trash detection problems. A well-organized dataset of sufficient size is essential to train object detection models and apply them to specific tasks. However, existing trash datasets have only a few hundred images, which are not sufficient to train deep neural networks. Most datasets are classification datasets that simply classify categories without location information. In addition, existing datasets differ from the actual guidelines for separating and discharging recyclables because the category definition is primarily the shape of the object. To address these issues, we build and experiment with trash datasets larger than conventional trash datasets and have more than twice the resolution. It was intended for general household goods. And annotated based on guidelines for separating and discharging recyclables from the Ministry of Environment. Our dataset has 10 categories, and around 33K objects were annotated for around 5K images with 1280×720 resolution. The dataset, as well as the pre-trained models, have been released at https://github.com/jms0923/tod.

• International Journal of Computer Science & Network Security
• /
• 제24권4호
• /
• pp.211-221
• /
• 2024

Pakistan is a top producer and exporter of high-quality rice, but traditional methods are still being used for detecting rice diseases. This research project developed an automated rice blast disease diagnosis technique based on deep learning, image processing, and transfer learning with pre-trained models such as Inception V3, VGG16, VGG19, and ResNet50. The modified connection skipping ResNet 50 had the highest accuracy of 99.16%, while the other models achieved 98.16%, 98.47%, and 98.56%, respectively. In addition, CNN and an ensemble model K-nearest neighbor were explored for disease prediction, and the study demonstrated superior performance and disease prediction using recommended web-app approaches.