• Journal of Broadcast Engineering
• /
• v.27 no.3
• /
• pp.341-350
• /
• 2022

In this paper, we propose a deep learning based on-device augmented reality (AR) system in which multiple input images are used to implement the correct occlusion in a real environment. The proposed system is composed of three technical steps; camera pose estimation, depth estimation, and object augmentation. Each step employs various mobile frameworks to optimize the processing on the on-device environment. Firstly, in the camera pose estimation stage, the massive computation involved in feature extraction is parallelized using OpenCL which is the GPU parallelization framework. Next, in depth estimation, monocular and multiple image-based depth image inference is accelerated using the mobile deep learning framework, i.e. TensorFlow Lite. Finally, object augmentation and occlusion handling are performed on the OpenGL ES mobile graphics framework. The proposed augmented reality system is implemented as an application in the Android environment. We evaluate the performance of the proposed system in terms of augmentation accuracy and the processing time in the mobile as well as PC environments.

• Journal of Korea Society of Industrial Information Systems
• /
• v.14 no.5
• /
• pp.139-148
• /
• 2009

In this paper, I propose an automatic analysis method of golf swine from single-camera video sequences. I define necessary swing features for automatic swing analysis in 2-dimensional environment and present efficient swing analysis methods using various image processing techniques including line and edge detection. The proposed method has two characteristics compared with previous swing analysis systems and related studies. First, the proposed method enables an automatic swing analysis in 2-dimension while previous systems require 3-dimensional environment which is relatively complex and expensive to run. Second, swing analysis is done automatically without human intervention while other 2-dimensional systems necessarily need analysis by a golf expert. I tested the method on 20 swing video sequences and found the proposed method works effective for automatic analysis of golf swing.

• Journal of the Korea Society of Computer and Information
• /
• v.14 no.3
• /
• pp.173-182
• /
• 2009

In this paper, we studied about the extraction of the parameter and implementation of speechreading system to recognize the Korean 8 vowel. Face features are detected by amplifying, reducing the image value and making a comparison between the image value which is represented for various value in various color space. The eyes position, the nose position, the inner boundary of lip, the outer boundary of upper lip and the outer line of the tooth is found to the feature and using the analysis the area of inner lip, the hight and width of inner lip, the outer line length of the tooth rate about a inner mouth area and the distance between the nose and outer boundary of upper lip are used for the parameter. 2400 data are gathered and analyzed. Based on this analysis, the neural net is constructed and the recognition experiments are performed. In the experiment, 5 normal persons were sampled. The observational error between samples was corrected using normalization method. The experiment show very encouraging result about the usefulness of the parameter.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2021.05a
• /
• pp.346-347
• /
• 2021

For pattern recognition for machine learning, the larger the amount of learning data, the better its performance. However, it is not always possible to secure a large amount of learning data with the types and information of patterns that must be detected in daily life. Therefore, it is necessary to significantly inflate a small data set for general machine learning. In this study, we study techniques to augment data so that machine learning can be performed. A representative method of performing machine learning using a small data set is the transfer learning technique. Transfer learning is a method of obtaining a result by performing basic learning with a general-purpose data set and then substituting the target data set into the final stage. In this study, a learning model trained with a general-purpose data set such as ImageNet is used as a feature extraction set using augmented data to detect a desired pattern.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2022.05a
• /
• pp.150-153
• /
• 2022

This paper reports a technique that automatically extracts object shapes through Dense-Net, and subsequently, detects the objects using Mask R-CNN in a manufacturing site, in which workers and objects are mixed. It is based on the customized factory dataset by targeting workers, machines, tools, control boxes, and products as the objects. Mask R-CNN supports multi-object recognition as a well-known object recognition method, while Dense-Net effectively extracts a feature from multiple and overlapping objects. After immediate implementation using the two technologies, the object is naturally extracted from a still image of the manufacturing site to describe image. Afterwards, the result is planned to be used to detect workers' abnormal behavior by adding a label on the objects.

• Transactions on Semiconductor Engineering
• /
• v.2 no.1
• /
• pp.17-20
• /
• 2024

This paper presents a novel lightweight object detection model tailored for low-powered edge devices, addressing the limitations of traditional resource-intensive computer vision models. Our proposed detector, inspired by the Single Shot Detector (SSD), employs a compact yet robust network design. Crucially, it integrates an 'enhancer block' that significantly boosts its efficiency in detecting smaller objects. The model comprises two primary components: the Light_Block for efficient feature extraction using Depth-wise and Pointwise Convolution layers, and the Enhancer_Block for enhanced detection of tiny objects. Trained from scratch on the Udacity Annotated Dataset with image dimensions of 300x480, our model eschews the need for pre-trained classification weights. Weighing only 5.5MB with approximately 0.43M parameters, our detector achieved a mean average precision (mAP) of 27.7% and processed at 140 FPS, outperforming conventional models in both precision and efficiency. This research underscores the potential of lightweight designs in advancing object detection for edge devices without compromising accuracy.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.24 no.2
• /
• pp.173-178
• /
• 2014

In this study, we introduce ASM-based face recognition classifier and its design methodology with the aid of 2-dimensional 2-directional hybird preprocessing algorithm. Since the image of face recognition is easily affected by external environments, ASM(active shape model) as image preprocessing algorithm is used to resolve such problem. In particular, ASM is used widely for the purpose of feature extraction for human face. After extracting face image area by using ASM, the dimensionality of the extracted face image data is reduced by using $(2D)^2$hybrid preprocessing algorithm based on LDA and PCA. Face image data through preprocessing algorithm is used as input data for the design of the proposed polynomials based radial basis function neural network. Unlike as the case in existing neural networks, the proposed pattern classifier has the characteristics of a robust neural network and it is also superior from the view point of predictive ability as well as ability to resolve the problem of multi-dimensionality. The essential design parameters (the number of row eigenvectors, column eigenvectors, and clusters, and fuzzification coefficient) of the classifier are optimized by means of ABC(artificial bee colony) algorithm. The performance of the proposed classifier is quantified through yale and AT&T dataset widely used in the face recognition.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.25 no.5
• /
• pp.502-508
• /
• 2015

Stereo images have an advantage of calculating depth(distance) values which can not analyze from 2-D images. However, depth information obtained by stereo images has due to following reasons: it can be obtained by computation process; mismatching occurs when stereo matching is processing in occlusion which has an effect on accuracy of calculating depth information. Also, if global method is used for stereo matching, it needs a lot of computation. Therefore, this paper proposes the method obtaining disparity map which can reduce computation time and has higher accuracy than established method. Edge extraction which is image segmentation based on feature is used for improving accuracy and reducing computation time. Color K-Means method which is image segmentation based on color estimates correlation of objects in an image. And it extracts region of interest for applying Loopy Belief Propagation(LBP). For this, disparity map can be compensated by considering correlation of objects in the image. And it can reduce computation time because of calculating region of interest not all pixels. As a result, disparity map has more accurate and the proposed method reduces computation time.

• Journal of the Korea Society of Computer and Information
• /
• v.14 no.3
• /
• pp.93-101
• /
• 2009

This paper suggests a new algorithm automatically searching for Region-of-Interest(ROI) with high speed, using the edge information of high frequency subband transformed with wavelet. The proposed method executes a searching algorithm of 4-direction object boundary by the unit of block using the edge information, and detects ROIs. The whole image is splitted by $64{\times}64$ or $32{\times}32$ sized blocks and the blocks can be ROI block or background block according to taking the edges or not. The 4-directions searche the image from the outside to the center and the algorithm uses a feature that the low-DOF image has some edges as one goes to center. After searching all the edges, the method regards the inner blocks of the edges as ROI, and makes the ROI masks and sends them to server. This is one of the dynamic ROI method. The existing methods have had some problems of complicated filtering and region merge, but this method improved considerably the problems. Also, it was possible to apply to an application requiring real-time processing caused by the process of the unit of block.

• KSCE Journal of Civil and Environmental Engineering Research
• /
• v.26 no.4D
• /
• pp.687-693
• /
• 2006

Image co-registration is the process of overlaying two images of the same scene, one of which represents a reference image, while the other is geometrically transformed to the one. In order to improve efficiency and effectiveness of the co-registration approach, the author proposed a pre-qualified area matching algorithm which is composed of feature extraction with canny operator and area matching algorithm with cross correlation coefficient. For refining matching points, outlier detection using studentized residual was used and iteratively removes outliers at the level of three standard deviation. Throughout the pre-qualification and the refining processes, the computation time was significantly improved and the registration accuracy is enhanced. A prototype of the proposed algorithm was implemented and the performance test of 3 Landsat images of Korea. showed: (1) average RMSE error of the approach was 0.435 pixel; (2) the average number of matching points was over 25,573; (3) the average processing time was 4.2 min per image with a regular workstation equipped with a 3 GHz Intel Pentium 4 CPU and 1 Gbytes Ram. The proposed approach achieved robustness, full automation, and time efficiency.