基于自适应脉冲耦合神经网络的水下激光图像分割方法  

Underwater Laser Image Segmentation Method based on Adaptive Pulse Coupled Neural Networks

作　　者：王博[1] 万磊[1] 李晔[1] 张铁栋[1] 

机构地区：[1]哈尔滨工程大学水下机器人技术重点实验室,黑龙江哈尔滨150001

出　　处：《光学学报》2015年第4期119-128,共10页Acta Optica Sinica

基　　金：国家自然科学基金(51009040);国家863计划(2011AA09A106)

摘　　要：距离选通式水下激光成像技术是一种能够有效抑制水介质的后向散射效应的探测技术,在海洋研究、深海探测和水下作业领域中拥有广阔的应用前景。然而在水下激光图像中出现的散斑噪声和灰度不均匀现象使得实现目标的准确分割较为困难。通过分析散斑噪声形成的机理,提出了一种水下激光图像的有效分割方法。该方法根据像素的噪声响应和灰度分布特性自适应确定各神经元的关键参数,并对噪声位置的神经元的行为进行抑制,基于最大二维Renyi熵准则采用梯度下降法确定了神经元的动态阈值,通过实验结果的比较分析说明该方法明显优于Normalized Cut、模糊C均值、均值漂移和分水岭分割方法,而运行时间约为常规脉冲耦合神经网络的五分之一。Range gated underwater laser imaging technology, which has broad application prospects in oceanic research, deep sea exploration and under water operation field, is one of the most effective methods to decrease the backward scattering effect of water medium. However, the special features of underwater laser images, such as speckle noise and non-uniform illumination, bring great difficulty for image segmentation. By analyzing the formation principle of speckle noise, an effective underwater laser image segmentation method is proposed. On the basis of noise response and intensity distribution, the proposed method determines the certain key parameters of neurons adaptively, while suppesses the behavior of neurons located in speckle noise. A gradient descent algorithm based on criterion of maximum two-dimensional Renyi entropy is applied to determine the dynamic threshold of neurons. Experimental results demonstrate that the method is significantly superior to Normalized Cut,fuzzy C means, mean shift and watershed methods, while the consumed time of this method is about one-fifth of conventional pulse coupled neural networks.

关 键 词：图像处理 水下激光图像分割 脉冲耦合神经网络 动态阈值 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]