• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2019.06a
• /
• pp.102-105
• /
• 2019

In the classical single-image super-resolution (SISR) reconstruction method using convolutional neural networks, the extracted features are not fully utilized, and the training time is too long. Aiming at the above problems, we proposed an improved SISR method based on a residual network. Our proposed method uses a feature fusion technology based on improved residual blocks. The advantage of this method is the ability to fully and effectively utilize the features extracted from the shallow layers. In addition, we can see that the feature fusion can adaptively preserve the information from current and previous residual blocks and stabilize the training for deeper network. And we use the global residual learning to make network training easier. The experimental results show that the proposed method gets better performance than classic reconstruction methods.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.07a
• /
• pp.540-543
• /
• 2020

기존의 초해상도 딥러닝 기법은 모델의 깊이가 깊어지면서, 좋은 성능을 내지만 점점 더 복잡해지고 있고, 실제로 사용하는데 있어 많은 시간을 요구한다. 이를 해결하기 위해, 우리는 딥러닝 모델의 가중치를 양자화 하여 추론시간을 줄이고자 한다. 초해상도 모델은 feature extraction, non-linear mapping, reconstruction 세 부분으로 나누어져 있으며, 레이어 사이에 많은 skip-connection 이 존재하는 특징이 있다. 따라서 양자화 시 최종 성능 하락에 미치는 영향력이 레이어 별로 다르며, 이를 감안하여 강화학습으로 레이어 별 최적 bit 를 찾아 성능 하락을 최소화한다. 본 논문에서는 Skip-connection 이 많이 존재하는 MSRN 을 사용하였으며, 결과에서 feature extraction, reconstruction 부분과 블록 내 특정 위치의 레이어가 항상 높은 bit 를 가짐을 알 수 있다. 기존에 영상 분류에 한정되어 사용되었던 혼합 bit 양자화를 사용하여 초해상도 딥러닝 기법의 모델 사이즈를 줄인 최초의 논문이며, 제안 방법은 모바일 등 제한된 환경에 적용 가능할 것으로 생각된다.

• Journal of Electrical Engineering and Technology
• /
• v.3 no.3
• /
• pp.436-443
• /
• 2008

This paper presents a new method for building detection and reconstruction from aerial images. In our approach, we extract useful building location information from the generated disparity map to segment the interested objects and consequently reduce unnecessary line segments extracted in the low level feature extraction step. Hypothesis selection is carried out by using an undirected graph, in which close cycles represent complete rooftops hypotheses. We test the proposed method with the synthetic images generated from Avenches dataset of Ascona aerial images. The experiment result shows that the extracted 3D line segments of the reconstructed buildings have an average error of 1.69m and our method can be efficiently used for the task of building detection and reconstruction from aerial images.

• Journal of IKEEE
• /
• v.12 no.1
• /
• pp.51-58
• /
• 2008

This paper presents a new method for building detection and reconstruction from aerial images. In our approach, we extract the useful building location information from the generated disparity map to obtain the segmentation of interested objects and thus reduce significantly unnecessary line segment extracted in low level feature extraction step. Hypothesis selection is carried out by using undirected graph in which close cycles represent complete rooftops hypotheses, and hypothesis are finally tested to contruct building model. We test the proposed method with synthetic images generated from Avenches dataset of Ascona aerial images. The experiment result shows that the extracted 3D line segments of the buildings can be efficiently used for the task of building detection and reconstruction from aerial images.

• ETRI Journal
• /
• v.42 no.3
• /
• pp.420-432
• /
• 2020

The reconstruction of archaeological fragments in 3D geometry is an important problem in pattern recognition and computer vision. Therefore, we implement an algorithm with the help of a 3D model to perform reconstruction from the real datasets using the slope features. This approach avoids the problem of gaps created through the loss of parts of the artifacts. Therefore, the aim of this study is to assemble the object without previous knowledge about the form of the original object. We utilize the edges of the fragments as an important feature in reconstructing the objects and apply multiple procedures to extract the 3D edge points. In order to assign the positions of the unknown parts that are supposed to match, the contour must be divided into four parts. Furthermore, to classify the fragments under reconstruction, we apply a backpropagation neural network. We test the algorithm on several models of ceramic fragments. It achieves highly accurate results in reconstructing the objects into their original forms, in spite of absent pieces.

• Journal of the Optical Society of Korea
• /
• v.15 no.1
• /
• pp.30-37
• /
• 2011

Binocular-pair images obtained from two cameras can be used to calculate the three-dimensional (3D) world coordinate of a feature point. However, to apply this method, measurement accuracy of binocular vision depends on some structure factors. This paper presents an experimental study of measurement distance, baseline distance, and baseline direction. Their effects on camera reconstruction accuracy are investigated. The testing set for the binocular model consists of a series of feature points in stereo-pair images and corresponding 3D world coordinates. This paper discusses a method to increase the baseline distance of two cameras for enhancing the accuracy of a binocular vision system. Moreover, there is an inflexion point of the value and distribution of measurement errors when the baseline distance is increased. The accuracy benefit from increasing the baseline distance is not obvious, since the baseline distance exceeds 1000 mm in this experiment. Furthermore, it is observed that the direction errors deduced from the set-up are lower when the main measurement direction is similar to the baseline direction.

• Proceedings of the Korean Society of Precision Engineering Conference
• /
• 2012.10a
• /
• pp.643-644
• /
• 2012

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.14 no.9
• /
• pp.3782-3796
• /
• 2020

A three-dimensional (3D) reconstruction is an important research area in computer vision. The ability to detect and match features across multiple views of a scene is a critical initial step. The tracking matrix W obtained from a 3D reconstruction can be applied to structure from motion (SFM) algorithms for 3D modeling. We often fail to generate an acceptable number of features when processing face or medical images because such images typically contain large homogeneous regions with minimal variation in intensity. In this study, we seek to locate sufficient matching points not only in general images but also in face and medical images, where it is difficult to determine the feature points. The algorithm is implemented on an adaptive threshold value, a scale invariant feature transform (SIFT), affine SIFT, speeded up robust features (SURF), and affine SURF. By applying the algorithm to face and general images and studying the geometric errors, we can achieve quasi-dense matching points that satisfy well-functioning geometric constraints. We also demonstrate a 3D reconstruction with a respectable performance by applying a column space fitting algorithm, which is an SFM algorithm.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.7 no.4
• /
• pp.839-844
• /
• 2003

To improve classification accuracy in this paper, we proposed an algorithm for the chromosome image reconstruction in the image preprocessing part. also we proposed the pattern classification method using the hierarchical multilayer neural network(HMNN) to classify the chromosome karyotype. It reconstructed chromosome images for twenty normal human chromosome by the image reconstruction algorithm. The four morphological and ten density feature parameters were extracted from the 920 reconstructed chromosome images. The each combined feature parameters of ten human chromosome images were used to learn HMNN(Hierarchical Multilayer Neural Network) and the rest of them were used to classify the chromosome images. The experimental results in this paper were composed to optimized HMNN and also obtained about 98.26% to recognition ratio.

• Korean Journal of Computational Design and Engineering
• /
• v.10 no.4
• /
• pp.262-273
• /
• 2005

For CAD data users, few things are as frustrating as receiving CAD data that is unusable due to poor data quality. Users waste time trying to get better data, fixing the data, or even rebuilding the data from scratch from paper drawings or other sources. Most related works and commercial tools handle the boundary representation (B-Rep) shape of CAD models. However, we propose a design history?based approach for healing CAD model errors. Because the design history, which covers the features, the history tree, the parameterization data and constraints, reflects the design intent, CAD model errors can be healed by an interdependency analysis of the feature commands or of the parametric data of each feature command, and by the reconstruction of these feature commands through the rule-based reasoning of an expert system. Unlike other B Rep correction methods, our method automatically heals parametric feature models without translating them to a B-Rep shape, and it also preserves engineering information.