Abstract: Because of the direct sunlight and shadow effect, it is difficult to segment cotton image in the natural environment. In order to segment cotton exactly, an improved segmentation algorithm based on watershed is proposed. In the method, a P-M diffusion pattern denoising filter is adopted in the original image. The multi-scale morphology gradient is modified by the hard threshold which gets from robust mean value estimation. Watershed transform is utilized to implement segmentation with the modifying gradient image and a region merging method based on L*a*b* space neighbor color similarity is employed to improve the segmentation result. The experiments show that the proposed algorithm has accurate segmentation result in the direct sunlight and shadow conditions.

Key words: cotton image segmentation, robust mean value estimation, neighbor color similarity, region merging

摘要： 针对棉花图像中存在阳光直射和阴影遮挡等因素而导致图像分割精度低、效果差的问题，提出一种改进分水岭的图像分割算法。该方法对原始图像进行各向异性扩散去噪预处理；利用鲁棒中值估计对形态学多尺度梯度图像进行硬阈值法梯度修正；对修正后的图像采用分水岭算法进行分割，对过分割的区域采用基于L*a*b*彩色空间的颜色相似度方案进行区域合并，从而将棉花提取出来。实验结果表明，提出的算法对阳光直射及阴影遮挡等干扰条件下的棉花图像分割能取得较好的效果。

关键词: 棉花图像分割, 鲁棒中值估计子, 颜色相似度, 区域合并