A Simple Image Coding by Projection of Principal Component in Segmented Color Areas

Abstract

This paper proposes a simple color image coding method using Principal Component Analysis (PCA) in the segmented color areas. A color image is segmented into different object areas with clustered color distributions. The chrominance a^* and b^* values in CIELAB space are observed to be strongly correlated with luminance L^* value in the object areas. After the segmentation, each object area is characterized by PCA. The segmented object areas are indexed by the class number which is greatly compressed by the conventional loss-less coding. The coded class number is transmitted with L^* image and the PCA parameters. PCA parameters are also compressed, because they are transmitted not by every pixel but by one set for each object area. The (a^*, b^*) values of each pixel are predicted by the projection of L* onto chromatic plane along to the first PC axis given by eigen vectors and are approximately restored from L^* value. L^* image can be compressed by the conventional lossive coding method such as JPEG or Wavelet. Finally, the full color image is reproduced by combining the luminance L^* with (a^*, b^*). The paper discusses the coding efficiency and the image quality changing with the class number.