• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.15 no.1
• /
• pp.180-194
• /
• 2021

For small image registration, feature-based approaches are likely to fail as feature detectors cannot detect enough feature points from low-resolution images. The classic FFT approach's prediction accuracy is high, but the registration time can be relatively long, about several seconds to register one image pair. To achieve real-time and high-precision rigid registration for small images, we apply deep neural networks for supervised rigid transformation prediction, which directly predicts the transformation parameters. We train deep registration models with rigidly transformed CIFAR-10 images and STL-10 images, and evaluate the generalization ability of deep registration models with transformed CIFAR-10 images, STL-10 images, and randomly generated images. Experimental results show that the deep registration models we propose can achieve comparable accuracy to the classic FFT approach for small CIFAR-10 images (32×32) and our LSTM registration model takes less than 1ms to register one pair of images. For moderate size STL-10 images (96×96), FFT significantly outperforms deep registration models in terms of accuracy but is also considerably slower. Our results suggest that deep registration models have competitive advantages over conventional approaches, at least for small images.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.46 no.1
• /
• pp.32-40
• /
• 2009

In this paper, we proposed an reconstruction algorithm of transmitted frames from displayed image on video terminal. For image quality assessment of the video streaming in the wireless network, we need information of the image that is transmitted to the end-user's device. Generally, subjective methods are widely used to evaluate the image quality by human beings because it is difficult to extract the transmitted image from the end-user's device. This paper presents an image reconstruction algerian based on the displayed image in video terminal for the extraction of the transmitted image. In the proposed method, we acquired the displayed image on video terminal using the camera. Camera-acquired images exhibit geometric and color distortions caused by characteristics of cameras and display devices. Therefore we correct the geometric distortion by exploiting the homography and color distortion by pre-computed look-up table. The experimental results show that the proposed measurement system yields promising estimation performance in terms of PSNR of $27{\sim}28dB$. We also carried out performance evaluation of the proposed method in terms of EPSNR and the quality of the estimated images by the proposed algerian was in fairly good range of MOS test scale.

• Journal of KIISE:Software and Applications
• /
• v.33 no.2
• /
• pp.252-266
• /
• 2006

Many problems in computer vision are mainly based on mathematical models. Their optimal solutions can be found by estimating the parameters of each model. However, provided an input data set is involved outliers which are relative]V larger than normal noises, they lead to incorrect results. RANSAC is a representative robust algorithm which is used to resolve the problem. One major problem with RANSAC is that it needs priori knowledge(i.e. a percentage of outliers) of the distribution of data. To solve this problem, we propose a FRANSAC algorithm which improves the rejection rate of outliers and the accuracy of solutions. This is peformed by categorizing all data into good sample set, bad sample set and vague sample set using a fuzzy classification at each iteration and sampling in only good sample set. In the experimental results, we show that the performance of the proposed algorithm when it is applied to the linear regression and the calculation of a homography.

• Journal of KIISE
• /
• v.42 no.10
• /
• pp.1185-1194
• /
• 2015

A spherical panoramic image does not conform to the pin-hole camera model, and, hence, it is not possible to utilize previous techniques consisting of plane-to-plane transformation. In this paper, we propose a new method to estimate the planar geometric transformation between the planar image and a spherical panoramic image. Our proposed method estimates the transformation parameters for latitude, longitude, rotation and scaling factors when the matching pairs between a spherical panoramic image and a planar image are given. A planar image is projected into a spherical panoramic image through two steps of nonlinear coordinate transformations, which makes it difficult to compute the geometric transformation. The advantage of using our method is that we can uncover each of the implicit factors as well as the overall transformation. The experiment results show that our proposed method can achieve estimation errors of around 1% and is not affected by deformation factors, such as the latitude and rotation.

• Journal of Broadcast Engineering
• /
• v.21 no.2
• /
• pp.200-209
• /
• 2016

This paper proposes a non-uniform de-blur algorithm using IMU sensor and a long/short exposure-time image pair to efficiently remove the blur phenomenon. Conventional blur kernel estimation algorithms using sensor information do not provide acceptable performance due to limitation of sensor performance. In order to overcome such a limitation, we present a kernel refinement step based on images having different exposure times which improves accuracy of the estimated kernel. Also, in order to figure out the phenomenon that conventional non-uniform de-blur algorithms suffer from severe degradation of visual quality in case of large blur kernels, this paper a homography-based residual de-convolution which can minimize quality degradation such as ringing artifacts during de-convolution. Experimental results show that the proposed algorithm is superior to the state-of-the-art methods in terms of subjective as well as objective visual quality.

• Journal of Digital Convergence
• /
• v.14 no.12
• /
• pp.231-236
• /
• 2016

Currently, a traffic monitoring service using a surveillance camera is provided through internet. In general, if the user points a certain location on a map, then this service shows the real-time image of the camera where it is mounted. In this paper, we proposed the intuitive surveillance monitoring system which displays a real-time camera image on the map by warping with bird's-eye view and with the top of image as the north. In order to robustly estimate the road plane using camera image, we used the motion vectors which can be detected to changes in brightness. We applied a re-adjustment process to have the same directivity with a map and presented a user-friendly interface that can be displayed on the map. In the experiment, the proposed method was presented as the result of warping image that the user can easily perceive like a map.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.23 no.4
• /
• pp.93-98
• /
• 2023

Recently, as the size of display devices tends to increase and taking pictures with smart phones has become commonplace, the need for taking high-resolution pictures with smart phones is increasing. However, when the lens size of a camera is limited, such as in a smartphone, there is a physical limit to increasing the resolution of a photo. This paper is about a technique for increasing the resolution of a picture even when using a small-sized lens like a smartphone camera. It is to take multiple pictures while moving the smartphone, and to increase the resolution by combining these pictures into one picture. First of all, two pictures were taken while moving the smartphone horizontally for the 2D picture. Processes such as camera matrix estimation, and homograph inverse transformation were performed using OpenCV, and the resolution was improved by synthesizing one picture. It was confirmed that the resolution was improved in parts such as oblique lines or arcs on several test pictures.