基于活跃度的脉冲耦合神经网络图像分割

Image segmentation based on activity degree with pulse coupled neural networks

针对脉冲耦合神经网络(PCNN)在图像分割中需多次人工调整网络参数的问题,提出了一种基于PCNN模型的图像自动分割方法。分析了图像中影响PCNN参数设置的因素,提出了一种图像自适应分块策略。将图像划分为内部复杂程度相近的子块,克服了同一参数无法同时对图像中复杂度差异较大的不同区域准确分割的不足。利用本文提出的局部图像活跃度(ADLA)指标对不同子块自适应地确定PCNN模型参数,有效解决了传统PCNN图像分割时需要人工确定关键参数的问题。最后,采用最大二维Tallis熵准则从分割后的二值结果序列中选择最佳结果。实验表明,本文方法的分割结果轮廓清晰、完整,即使在对比度低、背景呈大范围内连续变化等复杂情况下,也具有优异的性能。与传统PCNN分割方法相比,本文方法能自动、快速、准确地确定PCNN模型参数,且区域一致性测度(UM),区域对比度(CR),形状测度(SM),综合指标(CI)等客观评价的量化指标均优于传统PCNN分割方法12%以上。

As Pulse Coupled Neural Network（PCNN） in image segmentation has to adjust the network parameters by manual operation for several times, this paper proposed a automatic image segmentation method based on the PCNN. The effect factors on setting PCNN parameters in an image were ana- lyzed and an adaptive segment protocol was invited to divide the images into several sub-pieces with the similar inside complexities. By which, the weakness that the same parameter could not segment exactly the region with quite different complexities in an image at the same time was overcome. Fur- thermore, the index of Activity Degree of Local Area（ADLA） proposed by the paper was used to de- termine the PCNN model parameters for different sub-pieces adaptively and to avoid the manual opera- tion for important parameter selection in the traditional PCNN image segmentation, Finally the best result was chosen from the binary-result seaueneewith the max two-dimension Tallis entropy proto-col. The experiments proved that segmented figure is clear, complete and has excellent performance, even in conditions of a low contrast or a ehangeable background. Compared with the traditional PCNN segmentation methods, the proposed method not only can determine the PCNN model parameters au- tomatically, and its indexes on quantitative evaluation of the segmentation result, such as Uniformity Measure（UM）, Regional Contrast （CR）, Shape Measure （SM） and Comprehensive Index （CI） are all 12% better than those of the traditional PCNN method.