声纳图像成像质量差、特征信息弱,目标分割存在一定困难,为此提出一种融合多尺度统计信息的模糊C均值(FCM)聚类与Markov随机场(MRF)的小波域声纳图像分割算法.小波域中低频信息统计特性描述了低频不同区域像素聚类情况,高频信息反映了该方向纹理特征,依据低频子带的统计峰值选取FCM初始聚类中心,应用小波域FCM聚类算法对声纳图像进行预分割,抑制噪声的影响,提高了预分割的准确性;构建初分割后图像的多尺度MRF模型,尺度间节点标记的相关性采用1阶Markov性表征,尺度内构建2阶邻域系统描述系数间的标记联系,标记场采用双点多级逻辑模型建模,同一标记的系数特征场采用高斯模型建模,弥补了MRF算法中层次信息和轮廓信息描述的不足;应用迭代条件模型算法求其最小能量下的标记场,实现声纳图像分割.从视觉主观效果和客观评价指标两方面的实验结果验证表明,该算法分割声纳图像均优于FCM聚类算法和MRF算法,分割的声纳图像边缘与细节的清晰度、精细度均有一定程度改善.%Because of poor quality,weak feature information,and difficult target segmentation in sonar image,a sonar image segmentation algorithm based on fusing the fuzzy C-means (FCM) clustering of multi-scale statistical information and Markov random field (MRF) in the wavelet domain is proposed.In the wavelet domain,the low-frequency information depicts the clustering of pixels in different regions,and the high-frequency information reflects the texture feature in that direction.The proposed algorithm selects FCM initial cluster center from the low-frequency sub-band statistical peak value,and FCM algorithm is used for the pre-segmentation of sonar image and the suppression of the noise to improve the accuracy of pre-segmentation.The algorithm is used to construct multi-scale MRF model,and the correlation of inter-scale node marks is characterized by first order Markov.The intra-scale label connection among the description coefficients of two-order neighborhood system is constructed,double-point Multi-level Logistic (MLL) model is used in the label field,and Gauss model is used in the same-marked coefficient characteristic field,thus remeding the described shortfalls of hierarchical and silhouette information in MRF algorithm.The algorithm uses iteration condition model (ICM) algorithm to obtain its label fields for the minimum energy to realize the sonar image segmentation.The experimental results show that the proposed algorithm is better than FCM algorithm and MRF model algorithm,and the clarity and precision of the edge and details in the segmented sonar image are improved in a certain degree.
展开▼
