Atsushi Matsumura

A trainable retrieval system for cartoon character images

This paper proposes a novel method to retrieve cartoon character images in a database or network. In this method, partial features of an image, defined as regions and aspects, are used as keys to identify cartoon character images. The similarities between a query cartoon character image and the images in the database are computed by using these features. Based on the similarities, the cartoon images same or similar to the query image are identified and retrieved from the database. Moreover, our method adopts a training scheme to reflect the users subjectivity. The training emphasizes the significant regions or aspects by assigning more weight based on the users preferences and actions, such as selecting a desired image or an area of an image. These processes make the retrieval more effective and accurate. Experimental results verify the effectiveness and retrieval accuracy of the method.

Effective rate control method for minimizing temporal fluctuations in picture quality applicable for MPEG-4 AVC/H.264 encoding

Appropriate rate control plays a very important role in encoding motion pictures under the constant bit-rate. One of the requirements for rate control is minimizing temporal fluctuations in picture quality, which may cause flicker artifacts. To satisfy the requirement, some methods for MPEG-1 and MPEG-2 have been proposed. However, these methods cannot be applied to the H.264 encoding algorithm based on R-D optimization since they require the DCT coefficients and the accurate output bit amount for each quantization parameter of a current picture in advance. To overcome the problem, we propose a rate control method to satisfy the requirement in the H.264 encoding algorithm by introducing approximations to estimate the bit amount and the distortion.

A cartoon character retrieval system including trainable scheme

This paper proposes a novel method to retrieve cartoon character images in a database or network. In this method, partial features of a cartoon image, defined as regions and aspects, are used as keys to identify cartoon characters. The similarities between a query cartoon character image and the images in the database are computed by using these features. Based on the similarities, the cartoon images same or similar to the query image are identified and retrieved from the database. Moreover, our method adopts a training scheme to reflect the users subjectivity. The training emphasizes the significant regions or aspects based on the users preferences and actions, such as selecting a desired image or an area of an image. These processes make the cartoon-character retrieval more effective and accurate. Experimental results verify the effectiveness and retrieval accuracy of the method.

Effective interpolation for free viewpoint images using multilayered dynamic background buffers

Presenting images from a free viewpoint is a promising interactive video application. Though reference images from one viewpoint and their depth maps are often used to render free viewpoint images, picture quality degradation may occur because of lack of information in background regions that are occluded by foreground regions. In this paper an interpolation method for free viewpoint images using multilayered dynamic background buffers is proposed. In the proposed method, the buffers, updated using the reference images divided in each frame, are used to store the background regions and the output images are interpolated by the background buffers. Since the background buffers are created and updated using only the reference images and their depth maps, additional information on the background buffers is not required for the interpolation. The effectiveness of the proposed method was evaluated by several simulation experiments.