• International conference on construction engineering and project management
• /
• 2024.07a
• /
• pp.1284-1284
• /
• 2024

Leveraging large language models and safety accident report data has unique potential for analyzing construction accidents, including the classification of accident types, injured parts, and work processes, using unstructured free text accident scenarios. We previously proposed a novel approach that harnesses the power of fine-tuned Generative Pre-trained Transformer to classify 6 types of construction accidents (caught-in-between, cuts, falls, struck-by, trips, and other) with an accuracy of 82.33%. Furthermore, we proposed a novel methodology, saliency visualization, to discern which words are deemed important by black box models within a sentence associated with construction accidents. It helps understand how individual words in an input sentence affect the final output and seeks to make the model's prediction accuracy more understandable and interpretable for users. This involves deliberately altering the position of words within a sentence to reveal their specific roles in shaping the overall output. However, the validation of saliency visualization results remains insufficient and needs further analysis. In this context, this study aims to qualitatively validate the effectiveness of saliency visualization methods. In the exploration of saliency visualization, the elements with the highest importance scores were qualitatively validated against the construction accident risk factors (e.g., "the 4m pipe," "ear," "to extract staircase") emerging from Construction Safety Management's Integrated Information data scenarios provided by the Ministry of Land, Infrastructure, and Transport, Republic of Korea. Additionally, construction accident precursors (e.g., "grinding," "pipe," "slippery floor") identified from existing literature, which are early indicators or warning signs of potential accidents, were compared with the words with the highest importance scores of saliency visualization. We observed that the words from the saliency visualization are included in the pre-identified accident precursors and risk factors. This study highlights how employing saliency visualization enhances the interpretability of models based on large language processing, providing valuable insights into the underlying causes driving accident predictions.

• Journal of Sensor Science and Technology
• /
• v.10 no.1
• /
• pp.52-61
• /
• 2001

We propose a new saccadic eye movement system with visual selective attention. Saliency map models generate the scan pathways in a natural scene, of which the output makes an attended location. Saccadic eye movement model is used for producing the target trajectories to move the attended locations very rapidly. To categorize human saccadic eye movement, saccadic eye movement model was divided into three parts, each of which was then individually modeled using different neural networks to reflect a principal functionality of brain structures related with the saccadic eye movement in our brain. Based on the proposed saliency map models and the saccadic eye movement model, an active vision system using a CCD type camera and BLDC motor was developed and demonstrated with experimental results.

• The Transactions of the Korean Institute of Power Electronics
• /
• v.20 no.4
• /
• pp.305-312
• /
• 2015

In this study, inter-turn short fault models of interior permanent magnet synchronous motors (IPMSM) are developed by adding saliency modeling to surface-mounted permanent magnet motor models. The saliency model is obtained using the deformed flux models based on both fault-winding flux information and inductance variations caused by cross-flux linkages that depend on the distribution of the same phase windings. By assuming the balanced three-phase current injection, we obtain the positive and negative sequence voltages and the fault current in the positive and the negative synchronous reference frames. The output torque model is developed by adding the magnet and the reluctance torque, which are derived from the developed models. To verify the proposed IPMSM model with an inter-turn short fault, finite element method-based simulation and experimental measurement results are presented.

• Journal of the Institute of Electronics Engineers of Korea SC
• /
• v.49 no.1
• /
• pp.94-102
• /
• 2012

Recently entered into an aging socity as the number of elderly drivers is increasing. Traffic accidents of elderly drivers are caused by driver inattentions such as poor vehicle control due to aging, visual information retrieval problems caused by presbyopia, and objects identifying problems caused by low contrast sensitivity. In this paper, detection method of ROIs on the road is proposed. The proposed method creates the saliency map to detect the candidate ROIs from the input image. And, the input image is segmented to obtain the ROIs boundary. Finally, selective visual attention regions are detected according to the presence or absence of a segmented region with saliency pixels. Experimental results from a variety of outdoor environmental conditions, the proposed method presented a fast object detection and a high detection rate.

• Journal of Internet Computing and Services
• /
• v.12 no.6
• /
• pp.83-94
• /
• 2011

Main researching issue in affective computing is to give a machine the ability to recognize the emotion of a person and to react it properly. Efforts in that direction have mainly focused on facial and oral cues to get emotions. Postures have been recently considered as well. This paper aims to discriminate emotions posture by identifying and measuring the saliency of posture features that play a role in affective expression. To do so, affective postures from human subjects are first collected using a motion capture system, then emotional features in posture are described with spatial ones. Through standard statistical techniques, we verified that there is a statistically significant correlation between the emotion intended by the acting subjects, and the emotion perceived by the observers. Discriminant Analysis are used to build affective posture predictive models and to measure the saliency of the proposed set of posture features in discriminating between 6 basic emotional states. The evaluation of proposed features and models are performed using a correlation between actor-observer's postures set. Quantitative experimental results show that proposed set of features discriminates well between emotions, and also that built predictive models perform well.

• Proceedings of the KIPE Conference
• /
• 1998.10a
• /
• pp.630-635
• /
• 1998

This paper establishes in a new unified manner new mathematical models with core(iron) loss factor for two kinds of AC motors, induction and synchronous motors which are supposed to generate torque precisely or/and efficiently under vector controls. Our new models consist of three basic equations consistent with the others such as differential equation describing electromagnetic dynamics, torque equation describing torque generating mechanism, energy transmission equation describing how injected energy is wasted, saved or transmitted where all vector signals are defined in general frame of arbitrary instant angular velocity. It is clearly shown in our models that equivalent core-loss resistance can express appropriately and separately both eddy-current and hysteresis losses rather than mere vague loss. Proposed model of induction motor is the most compact in sense of the number of employed interior states and parameters. This compact model can also represent eddy-current and hysteresis losses of rotor as well as stator. For synchronous motor, saliency is taken into consideration. As well known model for cylindrical motor can be obtained directly from salient one as its special case.

• Journal of Broadcast Engineering
• /
• v.24 no.6
• /
• pp.1053-1063
• /
• 2019

Recently, tile-based streaming that transmits one 360 video in several tiles, is actively being studied in order to transmit these 360 video more efficiently. In this paper, for the transmission of high-definition 360 video corresponding to user's viewport in tile-based streaming scenarios, a system of assigning the quality of tiles at each tile by applying the saliency map generated by existing network models is proposed. As a result of usage of Motion-Constrained Tile Set (MCTS) technique to encode each tile independently, the user's viewport was rendered and tested based on Salient360! dataset, streaming 360 video based on the proposed system results in gain to 23% of the user's viewport compared to using the existing high-efficiency video coding (HEVC).

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.29 no.1
• /
• pp.117-125
• /
• 2019

Deep Learning, which is being used in various fields recently, is being threatened with Adversarial Attack. In this paper, we experimentally verify that the classification accuracy is lowered by adversarial samples generated by malicious attackers in image classification models. We used MNIST dataset and measured the detection accuracy by injecting adversarial samples into the Autoencoder classification model and the CNN (Convolution neural network) classification model, which are created using the Tensorflow library and the Pytorch library. Adversarial samples were generated by transforming MNIST test dataset with JSMA(Jacobian-based Saliency Map Attack) and FGSM(Fast Gradient Sign Method). When injected into the classification model, detection accuracy decreased by at least 21.82% up to 39.08%.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.16 no.4
• /
• pp.145-151
• /
• 2021

In this paper, we propose a novel visualization technique to explain the predictions of deep neural networks. We use knowledge distillation (KD) to identify the interior of a black-box model for which we know only inputs and outputs. The information of the black box model will be transferred to a white box model that we aim to create through the KD. The white box model will learn the representation of the black-box model. Second, the white-box model generates attention maps for each of its layers using Grad-CAM. Then we combine the attention maps of different layers using the pixel-wise summation to generate a final saliency map that contains information from all layers of the model. The experiments show that the proposed technique found important layers and explained which part of the input is important. Saliency maps generated by the proposed technique performed better than those of Grad-CAM in deletion game.

• Journal of the Institute of Electronics and Information Engineers
• /
• v.51 no.7
• /
• pp.157-168
• /
• 2014

We propose an novel artificial visual attention model that is capable of automatic detection and segmentation of saliency region on natural images in this paper. The proposed model is based on human visual perceptions in biological vision and contains there are main contributions. Firstly, we propose a novel framework of artificial visual attention model based on the opponent process theory using intensity and color features, and an entropy filter is designed to perceive salient regions considering the amount of information from intensity and color feature channels. The entropy filter is able to detect and segment salient regions in high segmentation accuracy and precision. Lastly, we also propose an adaptive combination method to generate a final saliency map. This method estimates scores about intensity and color conspicuous maps from each perception model and combines the conspicuous maps with weight derived from scores. In evaluation of saliency map by ROC analysis, the AUC of proposed model as 0.9256 approximately improved 15% whereas the AUC of previous state-of-the-art models as 0.7824. And in evaluation of salient region segmentation, the F-beta of proposed model as 0.7325 approximately improved 22% whereas the F-beta of previous state-of-the-art models.