• Journal of Broadcast Engineering
• /
• 제28권3호
• /
• pp.333-336
• /
• 2023

Recently, the transformer model used in natural language processing is also applied to the image super resolution field, showing good performance. However, these transformer based models have a disadvantage that they are difficult to use in small mobile devices because they are complex and have many learning parameters and require high hardware resources. Therefore, in this paper, we propose a knowledge distillation technique that can effectively reduce the size of a transformer based super resolution model. As a result of the experiment, it was confirmed that by applying the proposed technique to the student model with reduced number of transformer blocks, performance similar to or higher than that of the teacher model could be obtained.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 한국컴퓨터정보학회 2022년도 제66차 하계학술대회논문집 30권2호
• /
• pp.81-84
• /
• 2022

제조 산업에서의 이상치 검출은 생산품의 품질과 운영비용을 절감하기 위한 중요한 요소로 최근 딥러닝을 사용하여 자동화되고 있다. 이상치 검출을 위한 딥러닝 기법에는 CNN이 있으며, CNN을 계층적으로 구성할 경우 단일 CNN 모델에 비해 상대적으로 성능의 향상을 보일 수 있다는 것이 많은 선행 연구에서 나타났다. 이에 MVTec-AD 데이터셋을 이용하여 계층 CNN이 다중 클래스 이상치 판별 문제에 대해 효과적인지를 탐구하고자 하였다. 실험 결과 단일 CNN의 정확도는 0.7715, 계층 CNN의 정확도는 0.7838로 다중 클래스 이상치 판별 문제에 있어 계층 CNN 방식 접근이 다중 클래스 이상치 탐지 문제에서 알고리즘의 성능을 향상할 수 있음을 확인할 수 있었다. 계층 CNN은 모델과 파라미터의 개수와 리소스의 사용이 단일 CNN에 비하여 기하급수적으로 증가한다는 단점이 존재한다. 이에 계층 CNN의 장점을 유지하며 사용 리소스를 절약하고자 하였고 K-means, GMM, 계층적 클러스터링 알고리즘을 통해 제작한 새로운 클래스를 이용해 계층 CNN을 구성하여 각각 정확도 0.7930, 0.7891, 0.7936의 결과를 얻을 수 있었다. 이를 통해 Clustering 알고리즘을 사용하여 적절히 물체를 분류할 경우 물체에 따른 개별 상태 판단 모델을 제작하는 것과 비슷하거나 더 좋은 성능을 내며 리소스 사용을 줄일 수 있음을 확인할 수 있었다.

• The Journal of the Korea institute of electronic communication sciences
• /
• 제19권2호
• /
• pp.453-466
• /
• 2024

To improve the monitoring of Coastal Debris in the South Korea, which is difficult to estimate due to limited resources and vertex-based surveys, an approach based on UAV(Unmanned Aerial Vehicle) images and the RT-DETR(Realtime DEtection TRansformer) model was proposed for detecting Coastal Debris. By comparing to field investigation, the study suggested the possibility of quantitatively detecting coastal garbage and estimating the total capacity of garbage deposited on the natural coastline of the South Korea. The RT-DETR model achieved an accuracy of 0.894 for mAP@0.5 and 0.693 for mAP@0.5:0.95 in training. When applied to unmanaged coasts, the accuracy for the total number of coastal debris items was 72.9%. It is anticipated that if guidelines for defining monitoring of unmanaged coasts are established alongside this research, it should be possible to estimate the total capacity of the deposited coastal debris in the South Korea.

• The Journal of Bigdata
• /
• 제9권1호
• /
• pp.61-73
• /
• 2024

Class imbalance is one of the significant challenges in deep learning tasks, particularly pronounced in areas with limited data. This study proposes a new approach that utilizes minimal labeled data for effectively classifying tomato leaf diseases. We introduced a recursive learning method using the YOLOv8 model. By utilizing the detection predictions of images on the training data as additional training data, the number of labeled data is progressively increased. Unlike conventional data augmentation and up-down sampling techniques, this method seeks to fundamentally solve the class imbalance problem by maximizing the utility of actual data. Based on the secured labeled data, tomato leaves were extracted, and diseases were classified using the EfficientNet model. This process achieved a high accuracy of 98.92%. Notably, a 12.9% improvement compared to the baseline was observed in the detection of Late blight diseases, which has the least amount of data. This research presents a methodology that addresses data imbalance issues while offering high-precision disease classification, with the expectation of application to other crops.

• Journal of KIISE:Databases
• /
• 제37권2호
• /
• pp.71-81
• /
• 2010

The problem of finding clusters is widely used in numerous applications, such as pattern recognition, image analysis, market analysis. The important factors that decide cluster quality are the similarity measure and the number of attributes. Similarity measures should be defined with respect to the data types. Existing similarity measures are well applicable to numerical attribute values. However, those measures do not work well when the data is described by categorical attributes, that is, when no inherent similarity measure between values. In high dimensional spaces, conventional clustering algorithms tend to break down because of sparsity of data points. To overcome this difficulty, a subspace clustering approach has been proposed. It is based on the observation that different clusters may exist in different subspaces. In this paper, we propose a new similarity measure for clustering of high dimensional categorical data. The measure is defined based on the fact that a good clustering is one where each cluster should have certain information that can distinguish it with other clusters. We also try to capture on the attribute dependencies. This study is meaningful because there has been no method to use both of them. Experimental results on real datasets show clusters obtained by our proposed similarity measure are good enough with respect to clustering accuracy.

• Particle and aerosol research
• /
• 제13권4호
• /
• pp.151-158
• /
• 2017

Electrostatic precipitator is widely used to remove particulate matters in indoor air and industrial flue gas due to low pressure drop and high collection efficiency. However, it has a low collection efficiency for the submicrometer sized particles. Electrospraying is a potential method to increase the particle charging efficiency, which results in increased collection efficiency. Although particle charging efficiency is highly dependent upon droplet size, the effective measuring method of the droplets is still uncertain. Tap water was electrosprayed in this study, and the images of electrosprayed droplets were taken with a high speed camera coupled with several visualization methods in order to measure the droplets size. The droplet size distribution was determined by an image processing with an image-J program. As a result, a droplet measured by a laser visualization, had a half size of that by a Xenon light visualization. In addition, the experimentally measured droplet sizes were a good agreement with the predicted values suggested by $Fern{\acute{a}}ndez$ de la Mora and Loscertales(1994).

• KIPS Transactions on Software and Data Engineering
• /
• 제1권1호
• /
• pp.55-68
• /
• 2012

In this paper, we propose a novel method to solve the problem of excessive segmentation out of the method of segmenting regions from an image using Homogeneity Threshold($H_T$). The algorithm of the previous image segmentation based on $H_T$ was carried out region growth by using only the center pixel of selected window. Therefore it was caused resulting in excessive segmented regions. However, before carrying region growth, the proposed method first of all finds out whether the selected window is homogeneity or not. Subsequently, if the selected window is homogeneity it carries out region growth using the total pixels of selected window. But if the selected window is not homogeneity, it carries out region growth using only the center pixel of selected window. So, the method can reduce remarkably the number of excessive segmented regions of image segmentation based on $H_T$. In order to show the validity of the proposed method, we carried out multiple experiments to compare the proposed method with previous method in same environment and conditions. As the results, the proposed method can reduce the number of segmented regions above 40% and doesn't make any difference in the quality of visual image when we compare with previous method. Especially, when we compare the image united with regions of descending order by size of segmented regions in experimentation with the previous method, even though the united image has regions more than 1,000, we can't recognize what the image means. However, in the proposed method, even though image is united by segmented regions less than 10, we can recognize what the image is. For these reason, we expect that the proposed method will be utilized in various fields, such as the extraction of objects, the retrieval of informations from the image, research for anatomy, biology, image visualization, and animation and so on.

• Journal of Broadcast Engineering
• /
• 제24권5호
• /
• pp.713-725
• /
• 2019

Automatic License Plate Recognition (ALPR) is a technology required for many applications such as Intelligent Transportation Systems and Video Surveillance Systems. Most of the studies have studied were about the detection and recognition of license plates on cars, and there is very little about detecting and recognizing license plates on motorbikes. In the case of a car, the license plate is located at the front or rear center of the vehicle and is a straight or slightly sloped license plate. Also, the background of the license plate is mainly monochromatic, and license plate detection and recognition process is less complicated. However since the motorbike is parked by using a kickstand, it is inclined at various angles when parked, so the process of recognizing characters on the motorbike license plate is more complicated. In this paper, we have developed a 2-stage YOLOv2 algorithm to detect the area of a license plate after detection of a motorbike area in order to improve the recognition accuracy of license plate for motorbike data set parked at various angles. In order to increase the detection rate, the size and number of the anchor boxes were adjusted according to the characteristics of the motorbike and license plate. Image warping algorithms were applied after detecting tilted license plates. As a result of simulating the license plate character recognition process, the proposed method had the recognition rate of license plate of 80.23% compared to the recognition rate of the conventional method(YOLOv2 without image warping) of 47.74%. Therefore, the proposed method can increase the recognition of tilted motorbike license plate character by using the adjustment of anchor boxes and the image warping which fit the motorbike license plate.

• Tunnel and Underground Space
• /
• 제29권3호
• /
• pp.184-196
• /
• 2019

This study aims to extract a 3D image of micro-cracks generated by hydraulic fracturing tests, using the deep learning method and X-ray computed tomography images. The pixel-level cracks are difficult to be detected via conventional image processing methods, such as global thresholding, canny edge detection, and the region growing method. Thus, the convolutional neural network-based encoder-decoder network is adapted to extract and analyze the micro-crack quantitatively. The number of training data can be acquired by dividing, rotating, and flipping images and the optimum combination for the image augmentation method is verified. Application of the optimal image augmentation method shows enhanced performance for not only the validation dataset but also the test dataset. In addition, the influence of the original number of training data to the performance of the deep learning-based neural network is confirmed, and it leads to succeed the pixel-level crack detection.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• 제22권2호
• /
• pp.651-658
• /
• 2021

Because bolts with anti-loosening coatings are used mainly for joining safety-related components in automobiles, accurate automatic screening of these coatings is essential to detect defects efficiently. The performance of the convolutional neural network (CNN) used in a previous study [Identification of bolt coating defects using CNN and Grad-CAM] increased with increasing number of data for the analysis of image patterns and characteristics. On the other hand, obtaining the necessary amount of data for coated bolts is difficult, making training time-consuming. In this paper, resorting to the same VGG16 model as in a previous study, transfer learning was applied to decrease the training time and achieve the same or better accuracy with fewer data. The classifier was trained, considering the number of training data for this study and its similarity with ImageNet data. In conjunction with the fully connected layer, the highest accuracy was achieved (95%). To enhance the performance further, the last convolution layer and the classifier were fine-tuned, which resulted in a 2% increase in accuracy (97%). This shows that the learning time can be reduced by transfer learning and fine-tuning while maintaining a high screening accuracy.