Character recognition in scene images is being actively researched for the purpose of indexing in image retrieval. Character recognition in scene images is the technique of detecting and recognizing characters from general images taken with a digital camera. It needs high performance since most characters lie on backgrounds that have complicated textures and are geometrically distorted because of view angles. In this paper, we propose a novel method that starts by estimating the geometric distortion of the characters through support vector regression and then recognizes the character minus the distortion. Experiments show that the proposed method has higher recognition rate due to the distortion correction.

View full abstract

In this paper, a combination of random feature selection and bagging is proposed for achieving high accuracy in defect classification. First, elements and the number of features are selected randomly. Next, classification rates are estimated with bagging on various classifiers such as minimum distance classifier and support vector machine. Finally, the combination of features and classifier that achieve the best performance are found. Experimental results on real defect datasets show that our combination outperforms other approaches such as linear dimension reduction or Adaboost.

View full abstract

Automatic detection of intruder object and situation changes in scene images is an important research theme. This paper proposes a method for automatic detection of situation changes in cyclical moving scene-images based on the analysis of periodicity of movements. In the proposed method, cyclical movements are detected by voting binarized averaged grayscale values into each frequency. The occurrence of situation changes in a scene image can be detected by monitoring the change of bias conditions in voting. The experimental results showed that the speedy detections of situation changes were realized in the general scene images including cyclical moving objects.

View full abstract

We propose a non-photorealistic rendering (NPR) technique for generating an image of weaving patterns with color bamboo tapes. We use bamboo tapes of 8 colors on the vertecies of the RGB color cube. The bamboo-weave dithering method is devised for determining the color of horizontal and vertical tapes avoiding artificial repetitive color patterns and an iterative solution method is shown to be effective for reducing color quantization errors. We present an algorithm using iterative filtering for determining the weaving pattern of tapes, that is, which of horizontal or vertical tape passes above at their crossing points and show the algorithm is efficient for reducing the run-length of tape weaves. The height-field texture of weaves is obtained from the weaving pattern and is multiplied to dithered images for generating bamboo weave images.

View full abstract

We propose a method of constructing 3D building facade models using high resolution panoramic images. This system mounts three line scan cameras linked by a rotary encoder on a vehicle. The rotary encoder allows the method to capture each line image at regular intervals independent of vehicle speed. We can construct a 3D facade model by the stereo method. We mounted a prototype system and confirmed that high resolution panoramic images, 4mm/line, could be acquired while the vehicle was running at a speed of about 40km/h. 3D building facade models could be constructed from the high resolution panoramic images.

View full abstract

Except for television broadcasting about a big event etc., the information dispatch in the center from a local area is restricted. The Internet broadcasting is expected as media which can carry out information dispatch with a local capital at budget prices, without limiting the size and offer area of information. Then, in Saga Prefecture, it is experimenting in the local Internet broadcasting from 2004. In this paper, the design of the experiment system realized cheaply was described, and by the experiment over four years, while clarifying the problem of the local Internet broadcasting, realization possibility is suggested.

View full abstract

A non-photorealistic rendering technique is presented for morphing between two Voronoi-tessellated images. Objects in each image are firstly tessellated into Voronoi cells using the probabilistic point redistribution method and are then refined to centroidal Voronoi diagrams with the k-means algorithm. Every Voronoi cell in an image is continuously moved to its corresponding cell in another image.

View full abstract