We propose a prototype system that can be used by anyone to synthesize a caricature with not only the features of the target face but also the exaggeration he desires. First, we let 13 experts at drawing caricatures use our tool to synthesize line drawings corresponding to 5 targets' faces. Then, we learned their exaggeration rules from the synthesized line drawings by using neural networks. Our system traces the target face with a neutral facial expression and transforms it automatically into caricatures with various styles of exaggeration by using the following three types of exaggeration rules. The first type is the proposed exaggeration rule learned from the experts. The second is the traditional rule of simple extension of the geometric distance from an average face. The third is random deformation. Starting with one of these caricature types, a user can synthesize a caricature using an interactive genetic algorithm (IGA). We let 13 users synthesize their desired caricatures starting with caricatures generated only by exaggeration learned from experts and those generated only by extension of the geometric distance from an average face. Results of this experiment show that each type of exaggeration has advantages. These results also demonstrate the effectiveness of our system.
A non-photorealistic rendering technique has recently been presented for generating an image by weaving an input picture with horizontal and vertical bamboo strips. In this paper, we extend this technique for weaving two input images into both sides of a double-sided image without interfering each other. The input pictures are quantized and converted to reversible bamboo weave images with a similar pattern dither procedure.
Shaded surface model images like computer graphics (CG) need to be displayed for practical 3-D computer-generated holograms (CGH). We developed a simple process that is used to obtain 3-D data for CGH from two CG images. The first image is a shaded CG image. The second one is a depth-cued image, i.e., the image intensity is modulated by the depth of the object. We have implemented our method on OpenGL to render these two images. OpenGL includes support for not only removing the hidden surface but also texture mapping, alpha-blending, etc. We demonstrate the effectiveness of the method in achieving texture mapping and alpha-blending on the computer-generated rainbow holograms.
We developed additively weighted power centroidal Voronoi tessellation (APCVT) in which each generator of additively weighted power Voronoi tessellation coincides with the centroid of the corresponding Voronoi region and apply APCVT to non-photorealistic halftoning. The tone reproduction capability of the developed method is demonstrated on monochromatic and color images.
We have generated three dimensional wake waves by using computer graphics. A wake wave is a kind of wave made by things moving on or through water: objects such as ships, surfboards and animals like waterfowls or ducks. It is known that wake waves can be represented by combining Kelvin waves; however, there has only been studies on two dimensional representation and no study on three dimensional. Therefore, this paper describes how to generate three dimensional wake waves like those made by a waterfowl. To make three dimensional wake waves, the phase difference of Kelvin waves and the amplitude of wake waves are both necessary. In this study, we decided these parameters by measuring a real photograph of a waterfowl and by some ship references. Finally, we have generated three dimensional wake waves by using computer graphics and confirmed that the proposed method can generate wake waves similar to real ones by comparing the generated CG images with real photographs.
A method of controlling the viewing condition of each object in a 3D TV system consisting of a camera array and an integral photography display is presented. The method allows users to easily extract objects in the input multi-view images, and to independently control the positions of the extracted objects on the display by using different rendering parameters for different objects. It enables interactive editing of 3D scenes such as emphasizing the depth of a certain object and changing the relative positions of objects, rather than presenting the captured scene as it is.
We examined the technical requirements for the protection of private information held by new broadcasting services using communication networks. We surveyed organizations that collect and use private information as part of their service. One of their most important technical requirements is the protection of user information to avoid undermining user confidence in the organization. We have developed a database management system in response to this need. The system uses a secure distributed RDB instead of data encryption to protect private information. We evaluated our system by comparing its processing time against that of a conventional encryption method and demonstrated that it is practical in terms of processing time.
We developed a new pixel structure for a 3D real-time charge-coupled device imager. A background-level subtraction scheme and a charge integration cell enable us to measure 3D data under strong background light. In each pixel, the background level is extracted from the floating diffusion structure to extract the photoelectrons generated by the signal light without using complicated circuits. In addition, the integration cell reduces the photon shot noise generated by the background light. With these novel pixels, 3D real-time data can be detected even under sunlight of up to 100 k lx. .
We developed an MPEG-2 transcoding method on the basis of a two-tiered quantizer matrix that reduces re-quantization noise. The proposed method changes the quantization matrix to decrease the re-quantization noise, instead of changing the value of the quantization parameters as conventional methods do. Using this two-tiered matrix makes it possible to reduce re-quantization noise in the high-frequency domain. The main feature of the proposed method is that it does not cause any differences in image quality between adjacent macroblocks, which the conventional method does. Experimental results show that our proposed method produces transcoding scores of structural similarity (SSIM) and video quality model (VQM) that are 0.044 and 0.0938, respectively, better than those of the best conventional method. The proposed method always produced superior SSIM and VQM scores when the bit reduction ratio was smaller than 0.1015 for all the tested sequences.
We investigated how the arrangement of two packet scheduling algorithms (WFQ and FIFO) affects the application—level QoS of an audio—video streaming service over IP networks (typified by IPTV). We ran two computer simulations: (i) only the packet scheduling algorithms were applied, and (ii) both the packet scheduling algorithms and a media synchronization control were applied. Results showed that in case (i), a high application—level QoS with a relatively low delay was achieved when the WFQ was applied to congested nodes and to the node next to the media senders. In case (ii), a high QoS was achieved independently of the arrangement of WFQ nodes only if the bandwidth requirement was satisfied at the congested nodes.
To pinpoint a specific person from two or more individuals per image stored in a database, the order of importance of all individuals in the image must be analyzed. An algorithm is proposed that can be used to automatically analyze the order of importance on the basis of face size and the placement of a face within an image. The relative values among the major person and the others are also calculated. The proposed algorithm includes four steps: detection of faces, acquisition of composition information, selection of the major person, and judgment of order of importance. Tests on 100 images suggest that the proposed algorithm is suitable for analyzing the order of importance of individuals per image.