• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.11a
• /
• pp.510-512
• /
• 2022

컴퓨터 비전 분야에서 오랫동안 사용되었던 CNN(Convolution Neural Network)은 오분류를 일으키기 위해 악의적으로 추가된 섭동에 매우 취약하다. ViT(Vision Transformer)는 입력 이미지의 전체적인 특징을 탐색하는 어텐션 구조를 적용함으로 CNN의 국소적 특징 탐색보다 특성 픽셀에 섭동을 추가하는 적대적 공격에 강건한 특성을 보이지만 최근 어텐션 구조에 대한 강건성 분석과 다양한 공격 기법의 발달로 보안 취약성 문제가 제기되고 있다. 본 논문은 ViT가 CNN 대비 강건성을 가지는 구조적인 특징을 분석하는 연구와 어텐션 구조에 대한 최신 공격기법을 소개함으로 향후 등장할 ViT 파생 모델의 강건성을 유지하기 위해 중점적으로 다루어야 할 부분이 무엇인지 소개한다.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.23 no.3
• /
• pp.199-206
• /
• 2023

This study utilizes deep learning algorithms to automatically classify construction waste into three categories: wood waste, plastic waste, and concrete waste. Two models, VGG-16 and ViT (Vision Transformer), which are convolutional neural network image classification algorithms and NLP-based models that sequence images, respectively, were compared for their performance in classifying construction waste. Image data for construction waste was collected by crawling images from search engines worldwide, and 3,000 images, with 1,000 images for each category, were obtained by excluding images that were difficult to distinguish with the naked eye or that were duplicated and would interfere with the experiment. In addition, to improve the accuracy of the models, data augmentation was performed during training with a total of 30,000 images. Despite the unstructured nature of the collected image data, the experimental results showed that VGG-16 achieved an accuracy of 91.5%, and ViT achieved an accuracy of 92.7%. This seems to suggest the possibility of practical application in actual construction waste data management work. If object detection techniques or semantic segmentation techniques are utilized based on this study, more precise classification will be possible even within a single image, resulting in more accurate waste classification

• Proceedings of the Korea Information Processing Society Conference
• /
• 2021.11a
• /
• pp.749-752
• /
• 2021

오늘날 인공지능(Artificial Intelligence, AI) 분야에서 가장 구현하기 어려운 분야 중 하나는 추론이다. 근래 추론 분야에서 영상과 언어가 결합한 다중 모드(Multi-modal) 환경에서 영상 기반의 질의 응답(Visual Question Answering, VQA) 과업에 대한 AI 모델이 발표됐다. 얼마 지나지 않아 VQA 모델의 성능을 개선한 GVQA(Grounded Visual Question Answering) 모델도 발표됐다. 하지만 아직 GVQA 모델도 완벽한 성능을 내진 못한다. 본 논문에서는 GVQA 모델의 성능 개선을 위해 VCC(Visual Concept Classifier) 모델을 ViT-G(Vision Transformer-Giant)/14로 변경하고, ACP(Answer Cluster Predictor) 모델을 GPT(Generative Pretrained Transformer)-3으로 변경한다. 이와 같은 방법들은 성능을 개선하는 데 큰 도움이 될 수 있다고 사료된다.

• Smart Media Journal
• /
• v.12 no.9
• /
• pp.45-59
• /
• 2023

Weakly supervised object localization (WSOL) is a task of localizing an object in an image using only image-level labels. Previous studies have followed the conventional class activation mapping (CAM) pipeline. However, we reveal the current CAM approach suffers from problems which cause original CAM could not capture the complete defects features. This work utilizes a convolutional neural network (CNN) pretrained on image-level labels to generate class activation maps in a multi-scale manner to highlight discriminative regions. Additionally, a vision transformer (ViT) pretrained was treated to produce multi-head attention maps as an auxiliary detector. By integrating the CNN-based CAMs and attention maps, our approach localizes defective regions without requiring bounding box or pixel-level supervision during training. We evaluate our approach on a dataset of apple images with only image-level labels of defect categories. Experiments demonstrate our proposed method aligns with several Object Detection models performance, hold a promise for improving localization.

• Smart Structures and Systems
• /
• v.32 no.1
• /
• pp.23-35
• /
• 2023

Steel bolted joint is an important part of steel structure, and its damage directly affects the bearing capacity and durability of steel structure. Currently, the existing research mainly focuses on the identification of corroded bolts and corroded bolts respectively, and there are few studies on multiple states. A detection framework of corroded and loosened bolts is proposed in this study, and the innovations can be summarized as follows: (i) Vision Transformer (ViT) is introduced to replace the third and fourth C3 module of you-only-look-once version 5s (YOLOv5s) algorithm, which increases the attention weights of feature channels and the feature extraction capability. (ii) Three states of the steel bolts are considered, including corroded bolt, bolt missing and clean bolt. (iii) Line segment detector (LSD) is introduced for bolt rotation angle calculation, which realizes bolt looseness detection. The improved YOLOv5s model was validated on the dataset, and the mean average precision (mAP) was increased from 0.902 to 0.952. In terms of a lab-scale joint, the performance of the LSD algorithm and the Hough transform was compared from different perspective angles. The error value of bolt loosening angle of the LSD algorithm is controlled within 1.09%, less than 8.91% of the Hough transform. Furthermore, the proposed framework was applied to fullscale joints of a steel bridge in China. Synthetic images of loosened bolts were successfully identified and the multiple states were well detected. Therefore, the proposed framework can be alternative of monitoring steel bolted joints for management department.

• 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.