Abstract: Characteristics of different regions in side scan sonar imagery are analyzed. A kind of improved BEMD-multilayer level set segmentation algorithm is proposed to segment sonar imagery into three categories of region. In order to improve the accuracy of segmentation and the convergence rate of the level set model, BEMD（Two-dimensional Empirical Mode Decomposition） is used to describe the model's energy function. Through BEMD weighted parameters, the anti-noise performance of the model is improved without losing the accuracy of segmentation. After analysis of the contact of the c-means algorithm and the level set algorithm, an improved c-means algorithm is used to initialize the level set evolution curve. What’s more, a penalty term in the energy of level set is added to improve the evolution speed. Unsupervised sonar image segmentation experiments are done by using improved BEMD-hierarchical level set segmentation algorithm. Compared with other algorithms, the noise immunity, the segmentation accuracy and rapidity of the proposed algorithm are validated.

Key words: side scan sonar imagery, Two-dimensional Empirical Mode Decomposition（BEMD）, level set, energy function

摘要： 针对侧扫声纳图像不同区域的像素分布特点，提出了一种改进的BEMD（二维经验模态分解）-分层水平集分割算法。介绍了CV（Chan和Vese）水平集模型和分层水平集模型，利用分层水平集模型进行三类分割。为了提高分割精度，利用BEMD重新描述模型的能量函数。通过BEMD的加权参数，在不影响分割精度的前提下提高模型的抗噪性能。分析了c-均值算法与水平集算法的联系，利用改进的c-均值算法初始化水平集演化曲线，以减少迭代次数。对水平集能量函数添加惩罚项，以提高水平集演化速度。利用改进的BEMD-分层水平集分割算法进行无监督的图像分割实验并与其他算法比较，验证了该算法的抗噪性、分割的准确性和快速性。

关键词: 侧扫声纳图像, 二维经验模态分解, 水平集, 能量函数