This paper describes an immersive video transmission system that can provide users with a walk-through experience enabling them to feel as if they were passing through real objects. A server-client model is assumed for the transmission system configuration. Specifically, texture data and model data are generated and stored at the server in advance, and these signals are transmitted in response to the requested viewpoint. Finally, the view synthesis and the output image rendering are conducted at the client side. In the conventional scheme, the camera image is selected as the transmission content. However, this scheme requires a large number of camera images for precise reconstruction of occluded areas, and also requires a huge network bandwidth. To overcome this, we propose a new scheme in which the server stores orthogonal projection images as an alternative to texture data, and the appropriate selection from these images is realized. Our proposed scheme guarantees the successful reconstruction of so-called walk-through images under the basis of transmitting only two orthogonal projection images per object. The experimental results indicated that the proposed scheme was effective for reducing the amount of transmitted data compared to the conventional scheme, while maintaining the synthesized picture quality at an equivalent level.

抄録全体を表示

We propose a new coding method by applying slant transform (SLT) to gradation areas instead of discrete cosine transform (DCT). A contour in a gradation area, which is often caused by the coding process for high compression, is one of the typical artifact noises for subjective evaluation. Since the pixel value changes linearly in a gradation area, DCT is not always suitable for its transformation. In our proposed method, although only traditional DCT is utilized in the encoding process to avoid increasing computational complexity, the inverse SLT (ISLT) instead of IDCT is performed in the area that is estimated as a gradation area in accordance with the distribution of DCT coefficients in the decoding process. The coding block size of the gradation area is easily expanded by simple operations using the hierarchy of SLT. In accordance with our experimental analysis, the proposed scheme has been performed to clarify the improvement of quality in gradation areas.

抄録全体を表示

We propose a method for deblurring motion-blurred JPEG images that is based on an image restoration method. Motion-blurred JPEG images are degraded by the motion-blur as well as JPEG coding and decoding processes. The quantization noise on the JPEG coding, decoding, and the edge orientation are taken into consideration in this method. The results from our computer simulation prove the effectiveness of our method: while the images restored by the conventional restoration method are noisy, but the images deblurred using our proposed method are much better. Our method implicitly estimates the parameters by evaluating the deblurred image. The evaluation of the deblurred image is based on the skewness of the power spectrum of the output, which is obtained by applying a Measurement Differencing Filter to the deblurred image.

抄録全体を表示

We a proposed a novel stereo image editing technique based on a content-preserving image editing approach. This novel technique is for editing stereo images while preserving disparity without accurate estimation of disocclusion areas. In this paper, we describe a stereo extension of seam carving (SC) that is one of the content-preserving image processing operators. SC is used for image retargeting, which is to shrink or to enlarge size of an image, while still retaining the details of salient objects in the image. Traditional SC generates connected paths (seams) and eliminates them to reduce the size of the original image or duplicates them to enlarge. We introduce external energy functions that represent matching conditions for stereo images and find an optimized pair of seams from stereo images. Using the obtained stereo paired seams, we have demonstrated depth-preserving image retargeting for multi-view stereo images.

抄録全体を表示

Super-resolution image reconstruction is an important technology in many image processing areas, including image sensing, medical imaging, satellite imaging, and television signal conversion. It is also a key component of a recent consumer HDTV set that utilizes the CELL processor. Among various super-resolution methods, the learning-based method is one of the most promising. Its only problem is its enormous computational time for image searching from the large database of training images. We previously proposed a new total variation (TV) regularization super-resolution method that utilizes a learning-based super-resolution method and obtained excellent results in image quality improvement. However, this method requires a long computational time because of its use of the learning-based method. In the current study, we examine two methods of reducing this computational time. The proposed algorithms significantly reduce the complexity while maintaining a comparable image quality. This enables the application of learning-based super-resolution to motion pictures such as those on HDTV and Internet movies.

抄録全体を表示

In this paper, a novel adaptive loop filter model for video coding that predicts pixel values from local dithered binary patterns and a look-up table (LUT) is proposed. It can refine a decoded image adaptively to its local texture by using a LUT that is used in inverse halftoning techniques. The LUT is generated so that the sum of the improvements in the decoded image is maximized. Experimental results, in which the total bitrates including the LUT and PSNR were compared with the conventional method demonstrated the efficiency of the proposed method.

抄録全体を表示

In subband image coding, the optimum frequency band partition was previously proposed on the basis of the power spectrum of an input image. However, the partition pattern on the 2-D frequency domain was invariable when the bit rate changed. Therefore, a new optimum frequency band partition has been developed that can determine both the partition pattern on the 2-D frequency domain and the combination of quantizers with subband signals for the given bit rate. The experimental results show that the proposed method improved the coding performances at the lower bit rate compared with the conventional method.

抄録全体を表示

We propose a scheme for estimating the level of beat noise that adaptively cancels the noise from the received signals and thus solves a critical problem: minimizing the bit error rate of OCDMA systems. This scheme is implemented as an adaptive filter based on an expansion of the Volterra series: the received signal is used to estimate the key components of beat noise, transmitted signal and additive Gaussian noise. The Volterra series can be used to describe the simultaneous occurrence of optical signals from different transmitters, which causes the beat noise. We show the configuration, operation principles and performance of the adaptive filter. We also show that the proposed combination of maximum likelihood detection and the estimation of the beat noise reduces the bit error rate and thus raise the performance of a system so that it approaches the level offered by an equivalent system with no beat noise.

抄録全体を表示

We propose a method for generating maze-like images by connecting Voronoi regions using minimum spanning trees that link the generating points of the weighted centroidal Voronoi tessellations. We also propose a brightness correction method for the halftone image based on the hexagonal lattice approximation of the centroidal Voronoi tessellation. Experimental results showed that the obtained halftone images can reproduce the true continuous tone of input images.

抄録全体を表示

We propose a system that can capture the motion of an entire body from full high-definition (HD) television images. It is possible to precisely measure the movements of the body parts, i.e. the facial expression, trunk, arms, legs, and fingers using a high-resolution image. Processing high-resolution images requires a reduction of the computational efforts. We prepare a low-resolution image by downsizing an HD image to a quarter width and height, since capturing the motion of a person's trunk, arms, and legs does not always require a high-resolution image. After capturing the motion of the body from the low resolution image, we extract the images of the face and left and right hands from the HD image based on the captured motion so that each body part could be located in the center of the extracted images. The finger motion and facial expression are obtained from the HD quality images.

抄録全体を表示