In this paper, we propose the Active Net model and its application to region extraction. Active Net is a simulated two dimensional elastic network model which minimizes the energy functional and therefore shows both active and dynamic behavior. Its energy functional consists of the internal strain energy of the net and the image energy which attracts the net to features in a target region. The net deforms to wrap the region as the energy functional is minimized. We can apply Active Net to the extraction of a textured region. Some experimental results are also described.

View full abstract

An investigation into an interesting property of the Fourier spectral phase is presented. It is shown analytically that an inverse Fourier transform of only the spectral phase term in an image's spectrum produces the line contour image of the original. Also, the effect of the low bits coding of the phase is discussed from the analytical point of view.
Several application examples of these spectral phase properties to engineering technology are shown including a blurred image and a noisy image.

View full abstract

In computer vision, we need the significant edges that are boundaries between different surfaces, or an object and its background.
In this paper, we propose a method for detecting significant edges based on the local distribution of gradient magnitude values. The analysis enables us to tell whether any significant edges exist locally. In the edge area, the distribution of gradient magnitude values can apparently be separated into two clusters. In the non edge areas, the distribution seems to be one cluster. The significant edges in the edge area are detected by an automatic thresholding technique after which line thining takes place. Since the method does not use any smoothing criteria, detected edges are located where we desire. The experiments show that significant edges are detected and textural edges are suppressed.

View full abstract

This paper proposes an optimal histogram segmentation method using peak detection based on the consideration of statistical tests. For all peak detection algorithms proposed so far, results are dependent on the parameters to be specified. The method described, under the hypothesis of a mixture distribution formed by two neighboring peaks, tests whether to combine the peaks into one or keep them intact as two independent distributions according to the maximum location of squared Fisher distance. Test processing continues until no successive peaks can be combined. Peaks detected with different parameters converge to the same number, which can be considered as optimal, as it is verified that each peak can be treated as an independent distribution. In the experiment, the proposed method is applied to several gray-level histograms of images, and its effectiveness is verified by results. The optimal gray levels and thresholds for quantization can thus be estimated.

View full abstract

Relaxation neural network models have been studied to solve basic image processing problems such as quantization and intensity-surface interpolation. First, a relaxation neural network model is proposed to solve the multi-level representation problem for a gray-level image in local and parallel computations. This network iteratively minimizes the energy function defined by the local error in neighboring picture elements and generates high-quality multi-level images depending on local features. Next the applicability of the relaxation network model to intensity-surface interpolation of a gray-level image is studied from sparsely and irregularly sampled data. A relaxation network model is proposed to interpolate the missing gray levels in parallel, thus minimizing the energy function consisting of a membrane and thin plate. This method is effective for intensity-surface interpolation while preserving discontinuities of the image.

View full abstract

The Gradient-Based method is widely used to estimate motion parameters from images and provide information about the 3-Dimensional structure of rigid bodies.
To determine the optical flow from the Gradient-Based method, global or local optimization is used as an additional constraint. However, these methods are sensitive to noise and the optical flow obtained gets corrupted gradually around the boundaries of the moving regions.
This paper presents a method to detect optical flow using the intersections of brightness constraint equations. This method can detect the fine optical flow around the boundaries of the moving regions and proves to be robust to image noise.

View full abstract

This paper presents an edge-adaptive Kalman filter with ringing suppression for restoring motion-blurred images. The ringing artifacts are produced near the edges of images restored by conventional techniques of shift-invariant linear filtering. A one-dimensional non-stationary model is used as an original image, and the degradation model is represented by a uniform, shift-invariant, point-spread function and additive white noise. The image is divided into an edge region and a smooth region. Two Kalman filter gains are designed for these two regions. The image is then restored by the adaptive Kalman filter, which switches the two gains according to edge information. Model parameters and the edge information of the original image are estimated from the image temporarily restored by a deconvolution technique. Finally, computer simulations show that the proposed technique is effective for suppressing the ringing artifacts.

View full abstract

The paper describes an image-processing method to extract and track walking men from a series of frame images taken by an overhead video camera. First, each frame of image is divided into blocks of 4-by-4 pixels. Each block is then examined to determine whether it is part of a moving object or not. Four 8-by-8 pixels windows are set on the block under consideration to calculate the standard deviations of pixel values in each window. When these values of more than two windows vary according to the time change, the current block is regarded as a part of a walking man. Next, movement, appearance, dissappearance, connection and division of an object are automatically recognized from a series of frames. This method also can track a specified man. Algorithms of this method and experimental results are shown.

View full abstract

A new video synthesis method is proposed. In this method, the key for extracting moving objects is given in the first frame. In the following frames, the key is moved and transformed according to the detected motion-vector of the objects. Then with this transformed key, the objects are extracted automatically and synthesized. A motion-vector detection method suitable for this synthesis is discussed. According to the experiment using motion pictures with pan, tilt, zoom, and rotation, the matching method obtains an error smaller than ±2 pixels with detection points of 4 distant areas on the edge of the objects. It is confirmed that this synthesis method can be achieved at a video-rate on the real-time video signal processing system called the Picot-system.

View full abstract

The development of the Marathon image processing system which measures step frequency and automatically controls the camera by processing the image through the Marathon program is discussed. Rather than use a special image, this system processes the ordinary camera shot which can be broadcast. The image processes used for this purpose are bi-level quantize, area calculation, detection of a motion vector, chromakey, compress, expand and FFT. Some improvements were made to the motion vector process. These were implemented with the realtime image processing system-Picot system. Using the flexibility and generality of the Picot system, various processes have been developed very rapidly and effectively. Comparisons are drawn between general purpose software oriented system (Picot) and special purpose hardware oriented systems.

View full abstract

Since CG Simulation of natural terrain has applications in many areas, various modeling methods for terrain have been developed including imaginary forgeries. This paper presents a new approach suitable for modeling an imaginary mountain such as is produced by erosion processes. Such terrain can not be simulated by standard fractal surfaces. The algorithm synthesizes the forgery of natural mountain by refining the given initial ridge-lines which reflect the outline of the ideal forgery. The refinement is established by adding derivative ridge-lines recursively to those produced so far.
Although ridge-lines are the notable geographical feature of mountains produced by erosion processes, no suitable modeling method was designed previously.

View full abstract

A method of 3-D object reconstruction using Bézier patches with a subdivision scheme for recursive refinement of curved surface is described. The subdivision scheme is based first on using Bézier curves to approximate the boundaries of a surface patch, then on using Bézier patches to recursively approximate the curved surface. The advantage is that this method avoids the problem of ensuring the continuity of Bézier patches at the patch boundaries. A volume-approximation error metric is used to measure closeness of fit. The results of applying this method to shape 3-D object data from a slit-ray projection range system are presented. The advantage over the approximation method, which uses only planary patches, is demonstrated.

View full abstract

The method of image aquisition is one of the principal problems for handling super high definition pictures. This paper proposes a new method for acquiring high resolution pictures by processing two images taken with two different cameras. The method integrates two images into a higher resolution image. Compared to an imaging device with equivalent resolution, it will be less sensitive to the shot noise which becomes more dominant at the pixel size is reduced. Our theoretical analysis clearly shows the improvement in the high frequencies and details. The aperture effect determines the upper bound in the performance using this method. When the aperture ratio is 100%, the resolution improvement is limited to twice the original sampling frequency. Additionally, using more than two cameras, SNR of the reconstructed image improves even in the case of 100% aperture ratio. This means that the pixel size can be further reduced when extra cameras are used.

View full abstract

Many man-made objects such as industrial parts are partially constructed of surfaces of revolution, as well as planar surfaces. This paper studies the problem of finding surfaces of revolution in range data and proposes an approach based on the fact that at least one of two focal surfaces for a surface of revolution degenerates into the axis of rotation. For each point in the range image the centers of principal curvature, which construct the focal surface, are obtained from the first and second partial derivatives of the image and Hough transform is used to detect the axes of rotation by finding the centers which lie on straight lines in space. Finally, the revolution solid is volumetrically determined by estimating the radius function of cross section along each axis. The proposed method can be used where occlusion or truncation is a problem because it does not require the entire surfaces to be visible. The experiments have been successfully carried out using real range data.

View full abstract

In this paper, an efficient human image extraction method using a background reference image and binocular parallax are described. The objective is to extract a human image from general scenes which have no special backgrund. Images are acquired from a stereoscopic video camera system. A background reference image which does not include human image is acquired and stored in advance. Difference regions between input image and a stored backgrund refrence image are first detected. The human image region is then identified using binocular parallax. In the region detection process, those two images are compared pixel by pixel using intensity variance around the pixel and values of color compnents. This precise comparison and region detection make it easier to match the regions between left and right image.

View full abstract

In model-based coding of moving facial images, a 3-D-shape model of a head is prepared on both the transmitting and the receiving sides. The coding process is carried out by transmitting shape change and motion information for this shape model. “Luminance/chrominance” and “motion” information is two kinds of basic information treated in this coding method. This paper describes useful methods to represent “luminance/chrominance” and “motion” information so that it can be treated without restrictions on a specific shape model or input images. For “luminance/chrominance” information in each triangle composing the 3-D-shape model, the value is represented by 153 triangular lattice points. For “motion”, the detected shape change of each object is represented by object-dependent parameters and motion parameters. Experimental results show the effectiveness of the proposed methods.

View full abstract