When we use usual color spaces for image processing, there are some problems in the conventional image coding. In a usual color space each coordinate has correlation with another, and may represent unnecessary points because the shape of Gamut is complex. To solve the problems we propose Optimum Color Space Transform using principal component analysis. When the CIE_L*a*b* uniform color coordinates is transformed to the optimum color space, three new components have no correlation with one another. The optimum color space is efficient for image coding in terms of redundant reduction. Further more, this transform can be applied to color image analysis such as region segmentation, color restriction scheme. As a result, the optimum color space transform has a good performance for image coding and feature analysis.
We proposed a new region segmentation method for images containing text area and continuous-tone area in the same page. We calculate run-length of horizontal scanning using average density within the MxN pixel window containing a target pixel, judge the target pixel whether it makes an edge of a picture using the average density and maximum density defference, and count edges, calculate a rate of the edge pixels in a run, and classify each run to text area and continuous-tone area. As a result, our computer simulation shows that this method can work speedy and process line by line, and separate areas highly precise.
A new method to generate double scanning lines from two CCDs is described. Simply combining the outputs of normally driven CCDs With a vertical pixel shift can not generate a complete interlaced signal. We proposed a new 4 field sequence method to solve this problem. The result of the experiment showed that the vertical resolution of the two CCD system with the vertical pixel offset was almost doubled by this new method.
We have developed an excellent color reproducibility for single-chip CCD camera by using the mosaic filters of the complementary color filter. This process was performed by non linear polynomials including the terms of second degree and giving the compensationed operation to the luminance signal "Y" and the chrominance signals "R-Y" and "B-Y". The non linear polynomials were set to be optimized by taking the method of least squares applied to the color charts which Were chosen in equal hue, chroma, and value steps from the Munsell color standerds.
A photoelectronic GaAs integrated circuit, which converts light intensity to digital pulse frequency, has been made using the MES-FET process. The circuit includes a MSM (Metal-Semiconductor-Metal) photodiode (PD), a preamplifier (PA), a schmitt trigger (ST), a flip-flop (FF) (1 bit digital counter) and a reset (RS) FET. This sensor cell has wide dynamic range over 5 decades of incident light power and a γ characteristic suitable for a visible image device.
For parallel processing, we studied a new structure of a model "Holovision", which makes flexible recognition of visual patterns possible by means of an oscillating neural network. This structure is constituted by (1) plural number of elementary Holovisions each of which detects only two directions of lines, and (2) a supervising memory center which unites sets of symbols dispersed in the memory center of each elementary Holovision. Moreover, to realize stable recognition represented by coherence of non-linear oscillations, more suitable connections of neural units than those of previous model were applied to this model. The capability of this structure was demonstrated with computational experiments by using a geometrical pattern.