• Smart Structures and Systems
• /
• 제31권4호
• /
• pp.365-381
• /
• 2023

The existing vision-based techniques for inspection and condition assessment of civil infrastructure are mostly manual and consequently time-consuming, expensive, subjective, and risky. As a viable alternative, researchers in the past resorted to deep learning-based autonomous damage detection algorithms for expedited post-disaster reconnaissance of structures. Although a number of automatic damage detection algorithms have been proposed, the scarcity of labeled training data remains a major concern. To address this issue, this study proposed a semi-supervised learning (SSL) framework based on consistency regularization and cross-supervision. Image data from post-earthquake reconnaissance, that contains cracks, spalling, and exposed rebars are used to evaluate the proposed solution. Experiments are carried out under different data partition protocols, and it is shown that the proposed SSL method can make use of unlabeled images to enhance the segmentation performance when limited amount of ground truth labels are provided. This study also proposes DeepLab-AASPP and modified versions of U-Net++ based on channel-wise attention mechanism to better segment the components and damage areas from images of reinforced concrete buildings. The channel-wise attention mechanism can effectively improve the performance of the network by dynamically scaling the feature maps so that the networks can focus on more informative feature maps in the concatenation layer. The proposed DeepLab-AASPP achieves the best performance on component segmentation and damage state segmentation tasks with mIoU scores of 0.9850 and 0.7032, respectively. For crack, spalling, and rebar segmentation tasks, modified U-Net++ obtains the best performance with Igou scores (excluding the background pixels) of 0.5449, 0.9375, and 0.5018, respectively. The proposed architectures win the second place in IC-SHM2021 competition in all five tasks of Project 2.

• 한국컴퓨터정보학회논문지
• /
• 제29권7호
• /
• pp.11-20
• /
• 2024

본 논문에서는 생성 모델을 이용한 데이터 프리 양자화에서 발생할 수 있는 지식 격차를 줄이기 위하여 BAG (Bit-width Aware Generator)와 채널 어텐션 기반 중간 레이어 지식 증류를 제안한다. 생성 모델을 이용한 데이터 프리 양자화의 생성자는 오직 원본 네트워크의 피드백에만 의존하여 학습하기 때문에, 양자화된 네트워크의 낮은 bit-width로 인한 감소된 수용 능력 차이를 학습에 반영하지 못한다. 제안한 BAG는 양자화된 네트워크와 동일한 bit-width로 양자화하여, 양자화된 네트워크에 맞는 합성 이미지를 생성하여 이러한 문제를 완화한다. 또한, 양자화된 네트워크와 원본 모델 간의 지식 격차를 줄이는 것 역시 양자화에서 매우 중요한 문제이다. 이를 완화하기 위해 제안한 채널 어텐션 기반 중간 레이어 지식 증류는 학생 모델이 교사 모델로부터 어떤 채널에 더 집중해서 학습해야 하는지를 가르친다. 제안한 기법의 효율성을 보이기 위해, CIFAR-100에서 학습한 원본 네트워크를 가중치와 활성값을 각각 3-bit로 양자화하여 학습을 수행하였다. 그 결과 56.14%의 Top-1 Accuracy를 달성하였으며, 베이스라인 모델인 AdaDFQ 대비 3.4% 정확도를 향상했다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제15권12호
• /
• pp.4439-4455
• /
• 2021

Compared with two-stage object detection algorithms, one-stage algorithms provide a better trade-off between real-time performance and accuracy. However, these methods treat the intermediate features equally, which lacks the flexibility to emphasize meaningful information for classification and location. Besides, they ignore the interaction of contextual information from different scales, which is important for medium and small objects detection. To tackle these problems, we propose an image pyramid network based on dual attention mechanism (DAIPNet), which builds an image pyramid to enrich the spatial information while emphasizing multi-scale informative features based on dual attention mechanisms for one-stage object detection. Our framework utilizes a pre-trained backbone as standard detection network, where the designed image pyramid network (IPN) is used as auxiliary network to provide complementary information. Here, the dual attention mechanism is composed of the adaptive feature fusion module (AFFM) and the progressive attention fusion module (PAFM). AFFM is designed to automatically pay attention to the feature maps with different importance from the backbone and auxiliary network, while PAFM is utilized to adaptively learn the channel attentive information in the context transfer process. Furthermore, in the IPN, we build an image pyramid to extract scale-wise features from downsampled images of different scales, where the features are further fused at different states to enrich scale-wise information and learn more comprehensive feature representations. Experimental results are shown on MS COCO dataset. Our proposed detector with a 300 × 300 input achieves superior performance of 32.6% mAP on the MS COCO test-dev compared with state-of-the-art methods.

• 한국정보처리학회:학술대회논문집
• /
• 한국정보처리학회 2019년도 추계학술발표대회
• /
• pp.958-961
• /
• 2019

Convolutional neural network(CNN)의 등장으로 얼굴 표현 인식 연구는 많은 발전을 이루었다. 그러나, 기존의 CNN 접근법은 미리 학습된 훈련모델에서 Multiple-level 의 의미적 맥락을 포함하지 않는 Attention-embedded 문제가 발생한다. 사람의 얼굴 감정은 다양한 근육의 움직임과 결합에 기초하여 관찰되며, CNN 에서 딥 레이어의 산출물로 나온 특징들의 결합은 많은 서브샘플링 단계를 통해서 class 구별와 같은 의미 정보의 손실이 일어나기 때문에 전이 학습을 통한 올바른 훈련 모델 생성이 어렵다는 단점이 있다. 따라서, 본 논문은 Backbone 네트워크의 Multi-level 특성에서 Channel-wise Attention 통합 및 의미 정보를 포함하여 높은 인식 성능을 달성하는 Densely Backwarnd Attention(DBA) CNN 방법을 제안한다. 제안하는 기법은 High-level 기능에서 채널 간 시멘틱 정보를 활용하여 세분화된 시멘틱 정보를 Low-level 버전에서 다시 재조정한다. 그런 다음, 중요한 얼굴 표정의 묘사를 분명하게 포함시키기 위해서 multi-level 데이터를 통합하는 단계를 추가로 실행한다. 실험을 통해, 제안된 접근방법이 정확도 79.37%를 달성 하여 제안 기술이 효율성이 있음을 증명하였다.

• 방송공학회논문지
• /
• 제27권7호
• /
• pp.1011-1020
• /
• 2022

Recent high-performance panoptic segmentation models are based on transformer architectures. However, transformer-based panoptic segmentation methods are basically slower than convolution-based methods, since the attention mechanism in the transformer requires quadratic complexity w.r.t. image resolution. Also, sine and cosine computation for positional embedding in the transformer also yields a bottleneck for computation time. To address these problems, we adopt three modules to speed up the inference runtime of the transformer-based panoptic segmentation. First, we perform channel-level reduction using depth-wise separable convolution for inputs of the transformer decoder. Second, we replace sine and cosine-based positional encoding with convolution operations, called conv-embedding. We also apply a separable self-attention to the transformer encoder to lower quadratic complexity to linear one for numbers of image pixels. As result, the proposed model achieves 44% faster frame per second than baseline on ADE20K panoptic validation dataset, when we use all three modules.