前向后向扩散的距离正则模型应用于图像分割

摘要

This paper introduced a forward-and-backward diffusion-based distance regularized model for image segmentation to deal with the problem of irregularities that commonly appeared in level set evolution.It defined the distance regularization term (DRT)with a potential function such as the derived level set evolution had a unique forward-and-backward diffusion effect,i. e.,the diffusion was forward for steep shape region of the level set function(LSE),which kept decreasing the gradient magnitude until it approached 1 ,otherwise,the diffusion became backward and increased the gradient magnitude back to 1 .As a result,the LSE converge to sign distance function which was a desired shape of level set evolution.To demonstrate the effectiveness of the DRT,this paper applied it to an edge-based external energy for image segmentation.Experimental results show a good perform-ance of the distance regularized model,and the proposed method is robust for noisy and/or weak object images.%针对在演化过程中水平集函数振荡问题，提出前向后向扩散的距离正则模型应用于图像分割。新的距离正则项由一个势函数定义，推导的演化方程以唯一的方式前向后向扩散，即水平集函数在其陡峭区域前向扩散，降低函数的梯度模直至为1，反之它后向扩散，提高梯度模直至1。演化结果是水平集函数收敛于符号距离函数，这是水平集函数稳定演化所希望保持的状态；为了阐述距离正则项的有效性，将其与基于边缘信息的外能量项相结合。实验结果表明，该模型能够更好地完成图像分割，对噪声和弱目标图像鲁棒。