<title>New color segmentation method and its applications</title>

Abstract

Segmentation is an important step in the early stage of image analysis. Color or multi-spectral image segmentation usually involves search and clustering techniques in a three or higher dimensional spectral space - an exercise which is considered computationally expensive. This paper presents a new color segmentation method for color image analysis with its application to plant leaf area measurement. A 3D histogram for an RGB color image is established basing on an octree data structure. The histogram represents the color distribution of the image in the RGB color space on which a 3D Gaussian filter is applied to smooth out small maxima of this distribution. The color space is then searched to find out al the major maxima. Around each maxima, a covering cube with a controlled side width is established. These maxima and covering cubes are considered to be potential color classes. Each cube may expand according to the value of surrounding neighbors. Once enough modes and their cover cubes have been found, a k-means clustering algorithm is used to classify these maxima into a predetermined number of classes. Then, the classified modes and the color covered by the cubes are used as training samples for a Bayes classifier which can be used to classify all the pixels in the image. A statistical relaxation method is then sued as a find segmentation. This method can either be supervised or unsupervised, depending on the different requirements of specific applications. The octree data structure significantly reduces the color space to be searched and consequently reduces computational cost. An extension of this method can also be applied to multi-spectral image analysis.