[in Japanese]

View full abstract

Region growing is a widely used image-segmentation techniques. However, oversegmentation and dependency on the starting point (seed) pose difficulties. We propose au algorithm using the robust homogeneous criteria consisting of (1) intensity and (2) no edge in-between criterion. The region is grown into pixels that satisfy both criteria. The intensity criterion ensures that no abrupt. change occurs in the image intensity inside a region. It is calculated based on the local average image intensity after rejecting outliers. The no edge in-between criterion prevents the growing across the boundaries. The segmentation algorithm consists of a (1) coarse and a (2) fine scaled segmentation. In the first stage, large regions are segmented, which include both uniform regions and regions containing a small area of intensity fluctuation. In the second stage, the remaining regions are segmented. Our experiments showed good segmentation results for both gray-scale and color images.

View full abstract

This paper presents a method for designing a 6th-order two-channel linear-phase IIR filter bank with quasi-power complementary characteristics for image subband coding. The lowpass filter is specified to have a double or quadruple zero at ω = π and is designed with weighted equiripple approximation based on the formulation of an eigen-value problem. A design example of an image subband coding system using a three-stage octave division filter bank with rounded integer analysis output signals has shown to have a pseudo perfect reconstruction image “Lena” with a PSNR as highas 55.8 dB. The coding performance (PSNR versus bit rate) using bit plane arithmetic coding was nearly equal to that of 5th-order IIR filter banks and superior to 9 / 7 FIR filter banks.

View full abstract

The full-frame CCD image sensor is most suitable for digital still cameras. However, its disadvantages include low light sensitivity due to light absorption in the driving electrodes, a high dark current generated at the depleted Si-SiO₂ interface of the CCD register, and a small charge-handling capability due to the surface pinning mode of operation used to reduce dark current. In this paper, we propose a new CCD register that solves these problems. The cell of the register is an inverted version of the conventional photo-diode with overflow drains. Therefore, the register is suitable for back illumination. The overflow drains, which are isolated from each other with an SiO₂ layer, are used as driving electrodes. We analyze the performance of the register by using a three-dimensional numerical simulator and discuss its applicability to full-frame CCD image sensors. The leakage current between electrodes and between electrodes and channels was reduced to a negligible level by adjusting the barrier height. Also, the dark current generated at the Si-SiO₂ interface successfully flowed into the electrodes and decreased drastically. The transfer inefficiency for 6 × 10⁴ signal electrons, which is determined by the anti-blooming function of the cell, was as low as 10^-11 within a 8 ns transfer time for pushing a pulse with a 4 ns fall time. The new CCD register, which is driven by four-phase drive pulses, handled a charge amount that was several times larger than that of a two-phase CCD using surface pinning mode. We also discuss the relation between the register's design and its performance.

View full abstract

this paper proposes a new method for iteratively correcting three dimensional (3-d) shapes by weighting reconstruction errors when the shape is reconstructed from more than three 2-d perspective images using linear camera models such as weak-or para-perspective cameras. the method can compensate measurement errors by using its variance-covariance matrix, where the measurement errors are composed of random noise and the disparity between the linear camera models and non-linearity of perspective cameras, and can reduce mean reconstruction errors by adjusting the tolerance range of the reconstruction errors in accordance with the measurement error variances and also by eliminating the covariances between these errors. in numerical experiments using synthetic and real images, our method achieves correction results that are almost equivalent to those obtained by using variance-covariance matrices determined by a precise 3-d shape. the method also achieves a considerable amount of correction for both essentially nonlinear camera images and approximately linear camera images if the random noise is not strong and if more than twenty 2-d images are available.

View full abstract

[in Japanese]

View full abstract

[in Japanese]

View full abstract