• The Transactions of The Korean Institute of Electrical Engineers
• /
• v.68 no.2
• /
• pp.359-363
• /
• 2019

As commercial drones have been widely used, concerns for collision accidents with people and invading secured properties are emerging. The detection of drone is a challenging problem. The deep learning based object detection techniques for detecting drones have been applied, but limited to the specific cases such as detection of drones from bird and/or background. We have tried not only detection of drones, but classification of different drones with an end-to-end model. YOLOv2 is used as an object detection model. In order to supplement insufficient data by shooting drones, data augmentation from collected images is executed. Also transfer learning from ImageNet for YOLOv2 darknet framework is performed. The experimental results for drone detection with average IoU and recall are compared and analysed.

• Design & Manufacturing
• /
• v.15 no.3
• /
• pp.1-6
• /
• 2021

An artificial neural network model based on a deep learning algorithm is known to be more accurate than humans in image classification, but there is still a limit in the sense that there needs to be a lot of training data that can be called big data. Therefore, various techniques are being studied to build an artificial neural network model with high precision, even with small data. The transfer learning technique is assessed as an excellent alternative. As a result, the purpose of this study is to develop an artificial neural network system that can classify burr images of light guide plate products with 99% accuracy using transfer learning technique. Specifically, for the light guide plate product, 150 images of the normal product and the burr were taken at various angles, heights, positions, etc., respectively. Then, after the preprocessing of images such as thresholding and image augmentation, for a total of 3,300 images were generated. 2,970 images were separated for training, while the remaining 330 images were separated for model accuracy testing. For the transfer learning, a base model was developed using the NASNet-Large model that pre-trained 14 million ImageNet data. According to the final model accuracy test, the 99% accuracy in the image classification for training and test images was confirmed. Consequently, based on the results of this study, it is expected to help develop an integrated AI production management system by training not only the burr but also various defective images.

• ETRI Journal
• /
• v.46 no.4
• /
• pp.671-682
• /
• 2024

Violence can be committed anywhere, even in crowded places. It is hence necessary to monitor human activities for public safety. Surveillance cameras can monitor surrounding activities but require human assistance to continuously monitor every incident. Automatic violence detection is needed for early warning and fast response. However, such automation is still challenging because of low video resolution and blind spots. This paper uses ResNet50v2 and the gated recurrent unit (GRU) algorithm to detect violence in the Movies, Hockey, and Crowd video datasets. Spatial features were extracted from each frame sequence of the video using a pretrained model from ResNet50V2, which was then classified using the optimal trained model on the GRU architecture. The experimental results were then compared with wavelet feature extraction methods and classification models, such as the convolutional neural network and long short-term memory. The results show that the proposed combination of ResNet50V2 and GRU is robust and delivers the best performance in terms of accuracy, recall, precision, and F1-score. The use of ResNet50V2 for feature extraction can improve model performance.

• Journal of Information Processing Systems
• /
• v.17 no.3
• /
• pp.453-461
• /
• 2021

Many deep learning approaches have been studied for image classification in computer vision. However, there are not enough data to generate accurate models in medical fields, and many datasets are not annotated. This study presents a new method that can use both unlabeled and labeled data. The proposed method is applied to classify cervix images into normal versus cancerous, and we demonstrate the results. First, we use a patch self-supervised learning for training the global context of the image using an unlabeled image dataset. Second, we generate a classifier model by using the transferred knowledge from self-supervised learning. We also apply attention learning to capture the local features of the image. The combined method provides better performance than state-of-the-art approaches in accuracy and sensitivity.

• Smart Media Journal
• /
• v.9 no.4
• /
• pp.26-35
• /
• 2020

Attempts to predict future stock prices have been studied steadily since the past. However, unlike general time-series data, financial time-series data has various obstacles to making predictions such as non-stationarity, long-term dependence, and non-linearity. In addition, variables of a wide range of data have limitations in the selection by humans, and the model should be able to automatically extract variables well. In this paper, we propose a 'sliding time step normalization' method that can normalize non-stationary data and LSTM autoencoder to compress variables from all variables. and 'moving transfer learning', which divides periods and performs transfer learning. In addition, the experiment shows that the performance is superior when using as many variables as possible through the neural network rather than using only 100 major financial variables and by using 'sliding time step normalization' to normalize the non-stationarity of data in all sections, it is shown to be effective in improving performance. 'moving transfer learning' shows that it is effective in improving the performance in long test intervals by evaluating the performance of the model and performing transfer learning in the test interval for each step.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.24 no.11
• /
• pp.1403-1410
• /
• 2020

Recently, as interest in non-face-to-face experiences and services increases, the demand for web video contents that can be easily consumed using mobile devices such as smartphones or tablets is rapidly increasing. To cope with these requirements, in this paper we propose a technique to efficiently produce video contents that can provide experience of visiting famous places (i.e., stage tour) in animation or movies. To this end, an image dataset was established by collecting images of stage areas using Google Maps and Google Street View APIs. Afterwards, a deep learning-based style transfer method to apply the unique style of animation videos to the collected street view images and generate the video contents from the style-transferred images was presented. Finally, we showed that the proposed method could produce more interesting stage-tour video contents through various experiments.

• Journal of Internet Computing and Services
• /
• v.20 no.5
• /
• pp.1-8
• /
• 2019

This paper carried out a study to reduce damage from jellyfish whose population has increased due to global warming. The emergence of jellyfish on the beach could result in casualties from jellyfish stings and economic losses from closures. This paper confirmed from the preceding studies that the pattern of jellyfish's appearance is predictable through machine learning. This paper is an extension of The prediction model of emergence of Busan coastal jellyfish using SVM. In this paper, we used deep neural network to expand from the existing methods of predicting the existence of jellyfish to the classification by index. Due to the limitations of the small amount of data collected, the 84.57% prediction accuracy limit was sought to be resolved through data expansion using bootstraping. The expanded data showed about 7% higher performance than the original data, and about 6% better performance compared to the transfer learning. Finally, we used the test data to confirm the prediction performance of jellyfish appearance. As a result, although it has been confirmed that jellyfish emergence binary classification can be predicted with high accuracy, predictions through indexation have not produced meaningful results.

• Journal of the Institute of Convergence Signal Processing
• /
• v.24 no.1
• /
• pp.34-41
• /
• 2023

In recent years, as the range of human activities has increased, the introduction of alien species has become frequent. Among them, raccoons have been designated as harmful animals since 2020. Raccoons are similar in size and shape to raccoon dogs, so they generally need to be distinguished in capturing them. To solve this problem, we use VGG19, ResNet152V2, InceptionV3, InceptionResNet and NASNet, which are CNN deep learning models specialized for image classification. The parameters to be used for learning are pre-trained with a large amount of data, ImageNet. In order to classify the raccoon and raccoon dog datasets as outward features of animals, the image was converted to grayscale and brightness was normalized. Augmentation methods were applied using left and right inversion, rotation, scaling, and shift to create sufficient data for transfer learning. The FCL consists of 1 layer for the non-augmented dataset while 4 layers for the augmented dataset. Comparing the accuracy of various augmented datasets, the performance increased as more augmentation methods were applied.

• Journal of the Semiconductor & Display Technology
• /
• v.23 no.2
• /
• pp.135-142
• /
• 2024

Cu-Cu bonding, one of the key technologies in advanced packaging, enhances semiconductor chip performance, miniaturization, and energy efficiency by facilitating rapid data transfer and low power consumption. However, the quality of the interface bonding can significantly impact overall bond quality, necessitating strategies to quickly detect and classify in-process defects. This study presents a methodology for detecting defects in wafer junction areas from Scanning Acoustic Microscopy images using a ResNet-50 based deep learning model. Additionally, the use of the defect map is proposed to rapidly inspect and categorize defects occurring during the Cu-Cu bonding process, thereby improving yield and productivity in semiconductor manufacturing.

• Investigative Magnetic Resonance Imaging
• /
• v.25 no.1
• /
• pp.10-22
• /
• 2021

Purpose: To overcome the difficulty in building a large data set with a high-quality in medical imaging, a concept of 'blended-transfer learning' (BTL) using a combination of both source data and target data is proposed for the target task. Materials and Methods: Source and target tasks were defined as training of the source and target networks to reconstruct cardiac CINE images from undersampled data, respectively. In transfer learning (TL), the entire neural network (NN) or some parts of the NN after conducting a source task using an open data set was adopted in the target network as the initial network to improve the learning speed and the performance of the target task. Using BTL, an NN effectively learned the target data while preserving knowledge from the source data to the maximum extent possible. The ratio of the source data to the target data was reduced stepwise from 1 in the initial stage to 0 in the final stage. Results: NN that performed BTL showed an improved performance compared to those that performed TL or standalone learning (SL). Generalization of NN was also better achieved. The learning curve was evaluated using normalized mean square error (NMSE) of reconstructed images for both target data and source data. BTL reduced the learning time by 1.25 to 100 times and provided better image quality. Its NMSE was 3% to 8% lower than with SL. Conclusion: The NN that performed the proposed BTL showed the best performance in terms of learning speed and learning curve. It also showed the highest reconstructed-image quality with the lowest NMSE for the test data set. Thus, BTL is an effective way of learning for NNs in the medical-imaging domain where both quality and quantity of data are always limited.