• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.11a
• /
• pp.68-71
• /
• 2020

최근 딥러닝을 이용한 인체 자세 추정(human pose estimation) 연구가 활발히 진행되고 있다. 그 중 구조가 간단하면서도 성능이 강력하여 널리 사용되고 있는 딥러닝 네트워크 모델은 이미지 분류(image classification)에 사용되는 백본 네트워크(backbone network)와 디컨볼루션 출력망(deconvolution head network)을 이어 붙인 구조를 갖는다[1]. 기존의 디컨볼루션 출력망은 디컨볼루션 층을 쌓아 낮은 해상도의 특징맵을 모두 높은 해상도로 변환한 후 최종 인체 자세 추정을 하는데 이는 다양한 해상도에서 얻어낸 특징들을 골고루 활용하기 힘들다는 단점이 있다. 따라서 본 논문에서는 매 디컨볼루션 층 이후에 인체 자세 추정을 하여 다양한 해상도에서 연산을 하고 이를 종합하여 최종 인체 자세 추정을 하는 방법을 제안한다. 실험 결과 Res50 과 기존의 디컨볼루션 출력망의 경우 0.717 AP 를 얻었는데 Res101 과 기존의 디컨볼루션 출력망을 사용한 결과 50% 이상의 파라미터 수 증가와 함께 0.727 AP, 즉 0.010AP 의 성능 향상이 이루어졌다. 이에 반해 Res50 에 다중 해상도 디컨볼루션 출력망을 사용한 결과 약 1%의 파라미터 수 증가 만으로 0.720 AP, 즉 0.003 AP 의 성능 향상이 이루어졌다. 이를 통해 디컨볼루션 출력망 구조를 개선하면 매우 적은 파라미터 수 증가 만으로도 인체 자세 추정의 성능을 효과적으로 향상시킬 수 있음을 확인하였다.

• Journal of the Korea Convergence Society
• /
• v.11 no.12
• /
• pp.15-22
• /
• 2020

In this paper, we proposed a deep learning based super-resolution method that combines Channel Attention and Spatial Attention feature enhancement methods. It is important to restore high-frequency components, such as texture and features, that have large changes in surrounding pixels during super-resolution processing. We proposed a super-resolution method using feature enhancement that combines Channel Attention and Spatial Attention. The existing CNN (Convolutional Neural Network) based super-resolution method has difficulty in deep network learning and lacks emphasis on high frequency components, resulting in blurry contours and distortion. In order to solve the problem, we used an emphasis block that combines Channel Attention and Spatial Attention to which Skip Connection was applied, and a Residual Block. The emphasized feature map extracted by the method was extended through Sub-pixel Convolution to obtain the super resolution. As a result, about PSNR improved by 5%, SSIM improved by 3% compared with the conventional SRCNN, and by comparison with VDSR, about PSNR improved by 2% and SSIM improved by 1%.

• Journal of Digital Convergence
• /
• v.19 no.6
• /
• pp.251-258
• /
• 2021

In this paper, we proposed an encoder-decoder model utilizing residual learning to improve the accuracy of the U-Net-based semantic segmentation method. U-Net is a deep learning-based semantic segmentation method and is mainly used in applications such as autonomous vehicles and medical image analysis. The conventional U-Net occurs loss in feature compression process due to the shallow structure of the encoder. The loss of features causes a lack of context information necessary for classifying objects and has a problem of reducing segmentation accuracy. To improve this, The proposed method efficiently extracted context information through an encoder using residual learning, which is effective in preventing feature loss and gradient vanishing problems in the conventional U-Net. Furthermore, we reduced down-sampling operations in the encoder to reduce the loss of spatial information included in the feature maps. The proposed method showed an improved segmentation result of about 12% compared to the conventional U-Net in the Cityscapes dataset experiment.

• Journal of KIISE:Computer Systems and Theory
• /
• v.32 no.1
• /
• pp.9-20
• /
• 2005

Recently, the MPEG-7 standard a multimedia content description standard is wide]y used for content based image/video retrieval systems. However, since the descriptors standardized in MPEG-7 are usually multidimensional and the problem called 'Curse of dimensionality', previously proposed indexing methods(for example, multidimensional indexing methods, dimensionality reduction methods, filtering methods, and so on) could not be used to effectively index the multimedia database represented in MPEG-7. This paper proposes an efficient multimedia data indexing mechanism reflecting the characteristics of MPEG-7 visual descriptors. In the proposed indexing mechanism, the descriptor is transformed into a histogram of some attributes. By representing the value of each bin as a binary number, the histogram itself that is a visual descriptor for the object in multimedia database could be represented as a bit string. Bit strings for all objects in multimedia database are collected to form an index file, bitmap index, in the proposed indexing mechanism. By XORing them with the descriptors for query object, the candidate solutions for similarity search could be computed easily and they are checked again with query object to precisely compute the similarity with exact metric such as Ll-norm. These indexing and searching mechanisms are efficient because the filtering process is performed by simple bit-operation and it reduces the search space dramatically. Upon experimental results with more than 100,000 real images, the proposed indexing and searching mechanisms are about IS times faster than the sequential searching with more than 90% accuracy.

• Cartoon and Animation Studies
• /
• s.40
• /
• pp.157-176
• /
• 2015

This research describes render pass synthetic techniques required to use for the effectiveness of them in 3D animation synthetic technology. As the render pass is divided by property and synthesized after rendering, elaborate, rapid synthesis can be achieved. In particular, occlusion pass creates a screen as if it had a soft, light shading, expressing a sense of depth and boundary softness. It is converted into 2D image through a process of pass rendering of animation projects created in 3D space, then completed in synthetic software. Namely, 3D animation realizes the completeness of work originally planned through compositing, a synthetic process in the last half. To complete in-depth image, a scene manufactured in 3D software can be sent as a synthetic program by rendering the scene by layer and property. As recently the occlusion pass can express depth notwithstanding conducting GI rendering of 3D graphic outputs, it is an important synthetic map not omitted in the post-production process. Nonetheless, for the importance of it, currently the occlusion pass leaves much to be desired for research support and books summarizing and analyzing the characteristics of properties, and the principles and usages of them. Hence, this research was aimed to summarize the principles and usages of occlusion map, and analyze differences in the results of synthesis. Furthermore, it also summarized a process designating renderers and the map utilizing the properties, and synthetic software usages. For the future, it is hoped that effective and diverse latter expression techniques will be studied beyond the limitation of graphic expression based on trends diversifying technique development.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.5
• /
• pp.414-423
• /
• 2020

In this paper, we propose a Dilated Convolution Gate Linear Unit (DCGLU) to mitigate the lack of sparsity and small receptive field problems caused by the segmentation map extraction process in sound event detection with weak labels. In the advent of deep learning framework, segmentation map extraction approaches have shown improved performance in noisy environments. However, these methods are forced to maintain the size of the feature map to extract the segmentation map as the model would be constructed without a pooling operation. As a result, the performance of these methods is deteriorated with a lack of sparsity and a small receptive field. To mitigate these problems, we utilize GLU to control the flow of information and Dilated Convolutional Neural Networks (DCNNs) to increase the receptive field without additional learning parameters. For the performance evaluation, we employ a URBAN-SED and self-organized bird sound dataset. The relevant experiments show that our proposed DCGLU model outperforms over other baselines. In particular, our method is shown to exhibit robustness against nature sound noises with three Signal to Noise Ratio (SNR) levels (20 dB, 10 dB and 0 dB).

• The Korean Journal of Applied Statistics
• /
• v.35 no.3
• /
• pp.407-419
• /
• 2022

The semiconductor fabrication process is complex and time-consuming. There are sometimes errors in the process, which results in defective die on the wafer bin map (WBM). We can detect the faulty WBM by finding some patterns caused by dies. When one manually seeks the failure on WBM, it takes a long time due to the enormous number of WBMs. We suggest a two-step approach to discover the probable pattern on the WBMs in this paper. The first step is to separate the normal WBMs from the defective WBMs. We adapt a hierarchical clustering for de-noising, which nicely performs this work by wisely tuning the number of minimum points and the cutting height. Once declared as a faulty WBM, then it moves to the next step. In the second step, we classify the patterns among the defective WBMs. For this purpose, we extract features from the WBM. Then machine learning algorithm classifies the pattern. We use a real WBM data set (WM-811K) released by Taiwan semiconductor manufacturing company.

• Journal of the Korea Society of Computer and Information
• /
• v.27 no.8
• /
• pp.49-59
• /
• 2022

In this paper, we designed a new enzyme function prediction model PSCREM based on a study that compared and evaluated CNN and LSTM/GRU models, which are the most widely used deep learning models in the field of predicting functions and structures using protein sequences in 2020, under the same conditions. Sequence evolution information was used to preserve detailed patterns which would miss in CNN convolution, and the relationship information between amino acids with functional significance was extracted through overlapping RNNs. It was referenced to feature map production. The RNN family of algorithms used in small CNN-RNN models are LSTM algorithms and GRU algorithms, which are usually stacked two to three times over 100 units, but in this paper, small RNNs consisting of 10 and 20 units are overlapped. The model used the PSSM profile, which is transformed from protein sequence data. The experiment proved 86.4% the performance for the problem of predicting the main classes of enzyme number, and it was confirmed that the performance was 84.4% accurate up to the sub-sub classes of enzyme number. Thus, PSCREM better identifies unique patterns related to protein function through overlapped RNN, and Overlapped RNN is proposed as a novel methodology for protein function and structure prediction extraction.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.25 no.6
• /
• pp.819-824
• /
• 2021

As technologies related to image processing such as autonomous driving and CCTV develop, fog removal algorithms using a single image are being studied to improve the problem of image distortion. As a method of predicting fog density, there is a method of estimating the depth of an image by generating a depth map, and various fog features may be used as training data of the depth map. In addition, it is essential to implement a hardware capable of processing high-definition images in real time in order to apply the fog removal algorithm to actual technologies. In this paper, we implement NLCV (Normalize Local Coefficient of Variation), a feature of fog based on coefficient of variation, in hardware. The proposed hardware is an FPGA implementation of Xilinx's xczu7ev-2ffvc1156 as a target device. As a result of synthesis through the Vivado program, it has a maximum operating frequency of 479.616MHz and shows that real-time processing is possible in 4K UHD environment.

• Journal of the Korean Society for Precision Engineering
• /
• v.20 no.2
• /
• pp.176-183
• /
• 2003

The paper presents a feature recognition and segmentation method for surface approximation in reverse engineering. Efficient digitizing plays an important role in constructing a computational surface model from a physical part-surface without its CAD model on hand. Depending on its measuring source (e.g., touch probe or structured light), each digitizing method has its own strengths and weaknesses in terms of speed and accuracy. The final goal of the research focuses on an integration of two different digitizing methods: measuring by the structured light and that by the touch probe. Gathering bulk of digitized points (j.e., cloud-of-points) by use of a laser scanning system, we construct a coarse surface model directly from the cloud-of-points, followed by the segmentation process where we utilize the z-map filleting & differencing to trace out feature boundary curves. The feature boundary curves and the approximate surface model could be inputs to further digitizing by a scanning touch probe. Finally, more accurate measuring points within the boundary curves can be obtained to construct a finer surface model.