• Journal of the Korean Society of Radiology
• /
• v.18 no.1
• /
• pp.21-26
• /
• 2024

The one of the famous deep learning models for object detection task is you only look once version 5 (YOLOv5) framework based on the one stage architecture. In addition, YOLOv5 model indicated high performance for accurate lesion detection using the bottleneck CSP layer and skip connection function. The purpose of this study was to evaluate the performance of YOLOv5 framework according to various hyperparameters in position emission tomogrpahy (PET) phantom images. The dataset was obtained from QIN PET segmentation challenge in 500 slices. We set the bounding box to generate ground truth dataset using labelImg software. The hyperparameters for network train were applied by changing optimization function (SDG, Adam, and AdamW), activation function (SiLU, LeakyRelu, Mish, and Hardwish), and YOLOv5 model size (nano, small, large, and xlarge). The intersection over union (IOU) method was used for performance evaluation. As a results, the condition of outstanding performance is to apply AdamW, Hardwish, and nano size for optimization function, activation function and model version, respectively. In conclusion, we confirmed the usefulness of YOLOv5 network for object detection performance in nuclear medicine images.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.24 no.2
• /
• pp.93-98
• /
• 2024

With the rapid advance in AI (artificial intelligence) and high-performance computing technologies, deep learning is being used in various fields. Deep learning proceeds training by randomly reading a large amount of data and repeats this process. A large number of files are randomly repeatedly referenced during deep learning, which shows different access characteristics from traditional workloads with temporal locality. In order to cope with the difficulty in caching caused by deep learning, we propose a new sampling method that aims at reducing the randomness of dataset reading and adaptively operating on existing buffer cache algorithms. We show that the proposed policy reduces the miss rate of the buffer cache by 16% on average and up to 33% compared to the existing method, and improves the execution time by up to 24%.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.14 no.9
• /
• pp.3797-3822
• /
• 2020

The inspection of cracks on the surface of tunnel linings is a common method of evaluate the condition of the tunnel. In particular, determining the thickness and shape of a crack is important because it indicates the external forces applied to the tunnel and the current condition of the concrete structure. Recently, several automatic crack detection methods have been proposed to identify cracks using captured tunnel lining images. These methods apply an image-segmentation mechanism with well-annotated datasets. However, generating the ground truths requires many resources, and the small proportion of cracks in the images cause a class-imbalance problem. A weakly annotated dataset is generated to reduce resource consumption and avoid the class-imbalance problem. However, the use of the dataset results in a large number of false positives and requires post-processing for accurate crack detection. To overcome these issues, we propose a crack detection method using a ternary classifier. The proposed method significantly reduces the false positive rate, and the performance (as measured by the F1 score) is improved by 0.33 compared to previous methods. These results demonstrate the effectiveness of the proposed method.

• Journal of Institute of Control, Robotics and Systems
• /
• v.21 no.4
• /
• pp.367-371
• /
• 2015

In this paper, a novel localization method for a monocular camera is proposed by using a feature-based probabilistic map. The localization of a camera is generally estimated from 3D-to-2D correspondences between a 3D map and an image plane through the PnP algorithm. In the computer vision communities, an accurate 3D map is generated by optimization using a large number of image dataset for camera pose estimation. In robotics communities, a camera pose is estimated by probabilistic approaches with lack of feature. Thus, it needs an extra system because the camera system cannot estimate a full state of the robot pose. Therefore, we propose an accurate localization method for a monocular camera using a probabilistic approach in the case of an insufficient image dataset without any extra system. In our system, features from a probabilistic map are projected into an image plane using linear approximation. By minimizing Mahalanobis distance between the projected features from the probabilistic map and extracted features from a query image, the accurate pose of the monocular camera is estimated from an initial pose obtained by the PnP algorithm. The proposed algorithm is demonstrated through simulations in a 3D space.

• The Journal of Korea Robotics Society
• /
• v.15 no.2
• /
• pp.184-189
• /
• 2020

This paper is a study on data augmentation for small dataset by using deep learning. In case of training a deep learning model for recognition and classification of non-mainstream objects, there is a limit to obtaining a large amount of training data. Therefore, this paper proposes a data augmentation method using perspective transform and image synthesis. In addition, it is necessary to save the object area for all training data to detect the object area. Thus, we devised a way to augment the data and save object regions at the same time. To verify the performance of the augmented data using the proposed method, an experiment was conducted to compare classification accuracy with the augmented data by the traditional method, and transfer learning was used in model learning. As experimental results, the model trained using the proposed method showed higher accuracy than the model trained using the traditional method.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.9 no.4
• /
• pp.1424-1440
• /
• 2015

It is a challenging problem to search the intended images from a large number of candidates. Content based image retrieval (CBIR) is the most promising way to tackle this problem, where the most important topic is to measure the similarity of images so as to cover the variance of shape, color, pose, illumination etc. While previous works made significant progresses, their adaption ability to dataset is not fully explored. In this paper, we propose a similarity learning method on the basis of probabilistic generative model, i.e., probabilistic latent semantic analysis (PLSA). It first derives Fisher kernel, a function over the parameters and variables, based on PLSA. Then, the parameters are determined through simultaneously maximizing the log likelihood function of PLSA and the retrieval performance over the training dataset. The main advantages of this work are twofold: (1) deriving similarity measure based on PLSA which fully exploits the data distribution and Bayes inference; (2) learning model parameters by maximizing the fitting of model to data and the retrieval performance simultaneously. The proposed method (PLSA-FK) is empirically evaluated over three datasets, and the results exhibit promising performance.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.10 no.11
• /
• pp.5583-5604
• /
• 2016

In this paper, we present a novel image operator to extract textual information in natural scene images. First, a powerful refiner called the Stroke Color Extension, which extends the widely used Stroke Width Transform by incorporating color information of strokes, is proposed to achieve significantly enhanced performance on intra-character connection and non-character removal. Second, a character classifier is trained by using gradient features. The classifier not only eliminates non-character components but also remains a large number of characters. Third, an effective extractor called the Character Color Transform combines color information of characters and geometry features. It is used to extract potential characters which are not correctly extracted in previous steps. Fourth, a Convolutional Neural Network model is used to verify text candidates, improving the performance of text detection. The proposed technique is tested on two public datasets, i.e., ICDAR2011 dataset and ICDAR2013 dataset. The experimental results show that our approach achieves state-of-the-art performance.

• ETRI Journal
• /
• v.39 no.5
• /
• pp.643-651
• /
• 2017

Human age estimation is one of the key factors in the field of Human-Robot Interaction/Human-Computer Interaction (HRI/HCI). Owing to the development of deep-learning technologies, age recognition has recently been attempted. In general, however, deep learning techniques require a large-scale database, and for age learning with variations, a conventional database is insufficient. For this reason, we propose an age estimation method using artificially generated data. Image data are artificially generated through 3D information, thus solving the problem of shortage of training data, and helping with the training of the deep-learning technique. Augmentation using 3D has advantages over 2D because it creates new images with more information. We use a deep architecture as a pre-trained model, and improve the estimation capacity using artificially augmented training images. The deep architecture can outperform traditional estimation methods, and the improved method showed increased reliability. We have achieved state-of-the-art performance using the proposed method in the Morph-II dataset and have proven that the proposed method can be used effectively using the Adience dataset.

• KIPS Transactions on Software and Data Engineering
• /
• v.5 no.11
• /
• pp.569-584
• /
• 2016

In this paper, we present a spatial knowledge extractor implemented in Hadoop MapReduce parallel, distributed computing environment. From a large spatial dataset, this knowledge extractor automatically derives a qualitative spatial knowledge base, which consists of both topological and directional relations on pairs of two spatial objects. By using R-tree index and range queries over a distributed spatial data file on HDFS, the MapReduce-enabled spatial knowledge extractor, MRSPAKE, can produce a web-scale spatial knowledge base in highly efficient way. In experiments with the well-known open spatial dataset, Open Street Map (OSM), the proposed web-scale spatial knowledge extractor, MRSPAKE, showed high performance and scalability.

• Nuclear Engineering and Technology
• /
• v.52 no.5
• /
• pp.1008-1012
• /
• 2020

In order to make reasonable design for the improvement of comprehensive mechanical properties of RAFM steels, the design system with both machine learning and high-throughput optimization algorithm was established. As the basis of the design system, a dataset of RAFM steels was compiled from previous literatures. Then, feature engineering guided random forests regressors were trained by the dataset and NSGA II algorithm were used for the selection of the optimal solutions from the large-scale solution set with nine composition features and two treatment processing features. The selected optimal solutions by this design system showed prospective mechanical properties, which was also consistent with the physical metallurgy theory. This efficiency design mode could give the enlightenment for the design of other metal structural materials with the requirement of multi-properties.