Yong-In Yoon

An efficient method to build panoramic image mosaics

This paper describes an efficient method to build panoramic image mosaics with multiple images. Conventional algorithms used geometrical feature points and optimization to compute the projective transformation, which is the relation between two consecutive images. However, building a panoramic image was very time consuming because of the iterative computation involved.The proposed method computed the projective transformation in overlapped areas of the two given images by using four seed points. The seed point is the highly textured point in the overlapped area of the reference image, which is extracted by using phase correlation. Because the region of interest (ROI) was restricted within overlapped areas of two images, more accurate correspondences were obtained. Before selecting the seed point, the histograms of the overlapped areas were equalized to mitigate the variation of the illumination conditions. After selecting the seed point, the weighted block matching algorithm (BMA) was used to minimize image distortion caused by camera rotation. An experiment was performed employing the proposed method with various images and the results were compared with peak signal to noise ratio (PSNR). Results showed that the proposed method built high-quality panoramic image mosaics in high speed.

Facial component detection for efficient facial characteristic point extraction

This paper proposes an algorithm detecting facial component to efficiently extract the FCP (Facial Characteristic Point). The FCP plays an important role in facial expression representation or recognition. For efficient FCP extraction using image processing, we analyze and improve the conventional algorithms detecting facial components that are the basis of the FCP extraction. The proposed algorithm includes face region detection without the effect of skin-color hair, eye region detection with weighted template, eyebrow region detection using a modified histogram, and mouth region detection using skin characteristics.

AR Squash Game

In this paper, we describe AR squash game, which is a squash-style game based on AR (augmented reality) technique. We made this game using AR technique to estimate geometric information of images taken from a CCD camera. A CCD camera takes real-time images of a room formed by some number of walls, and the geometric information of the image is estimated using markers attached on the walls. By the geometric information and motion data of a ball created internally, a realistic scene is generated, in which the ball moves like a real one. And a user watches the computer screen displaying the mixed image, and tosses the ball using a racket connected to motion tracker. The position and the motion data of the racket is received from a 3D motion tracker

Detecting digital forgeries using of a lighting direction

This paper proposes a new method for digital forensic using a technique to detect a lighting direction. We describe how the estimated the lighting direction from a segmented image and show a ratio of detecting forgery.

Motion vector recovery by surrounding region matching based on gradient difference

The error concealment is a very useful technique for real-time communication systems, such as video conferencing. In this paper, we propose an error concealment technique designated to minimize discontinuities between block boundaries in consideration of the fact that the human visual system is sensitive to discontinuities. Error concealment for each loss block is performed by surrounding region matching of the loss block based on gradient difference. By computer simulation, the performance improvement for the proposed technique is shown to be superior than 1.32 dB in terms of the PSNR, in comparison with the conventional techniques.

3D-based synthesis and 3D reconstruction from uncalibrated images

In this paper, we propose a new camera calibration method for 3D-based image synthesis and 3D reconstruction. We improve the problem by changing the principle point for obtaining the linear equation. According to the error rate, we adapt the non-linear method that minimizes the intrinsic parameters. Namely, it minimizes the intrinsic parameters error by maintaining the computational conciseness. As a result, we can find optimized camera intrinsic parameters and adapt them for image synthesis and reconstruction.

Reconstruction of linearly parameterized models from a single image using the vanishing points

We present a method using only three vanishing points to recover the dimensions of object and its pose from single image of perspective projection with a camera of unknown focal length. Our approach is to compute the dimensions of objects represented by the unit vector of objects from the image. The dimension vector v of objects can be solved by the standard nonlinear optimization techniques with a multistart method which generates multiple starting points for the optimizer by sampling the parameter space uniformly. This method allows model-based vision to be computed the dimensions of object for a 3D model from matches to a single 2D image. Experimental results demonstrate the dimension vector v of the proposed method from a single image using three vanishing points and show a performance of the proposed method compared to the conventional. Then, the actual dimensions of object from the image agree well with the calculated results.

Reconstruction of linearly parameterized models three vanishing points from a single image of perspective projection

A method using only three vanishing points to recover the dimensions of object and its pose from a single image of perspective projection with a camera of unknown focal length is presented. Our approach is to compute the dimensions of objects which are represented by the unit vector of objects from the image. The dimension vector v for the objects can be solved by the standard nonlinear optimization techniques with a multistart method which generates multiple starting points for the optimizer by sampling the parameter space uniformly. This method allows model-based vision to be computed the dimensions of object for a 3D model from matches to a single 2D image. Therefore, experimental results show the evaluated values of the object dimensions from a single image using three vanishing points. In addition, the actual dimensions of object from the image agree well with the calculated results.