Based On The Histogram And Pulse Coupled Neural Network And The Marginal Product Of The Mutual Information Of Image Segmentation

Image segmentation refers to the process of partitioning a digital image into interest multiple disjoint segments,and each region would meet the consistency or similarity of some features in a given area. It is regarded as one the most difficult international academic problems in image understanding and computer vision field for a long time. So far, although there exists a variety of segmentation methods, yet no general segmentation algorithm suits all images. Common segmentation methods are edge detection based algorithm, threshold method, region growing method, coordinate mapping method,neural networks method and so on. Neural network is a computing architecture simulating biological process and reflecting some biological neural characteristics.In particular, image segmentation based on pulse coupled neural networks totally depends on natural attributes of images, without pre-selecting the spatial extend and training or the choice of parameters. Relative to the classical active contour models, level set, clustering algorithms and other advanced methods, unit-linking pulse coupled neural network could guarantee the accuracy of segmentation and efficiency advantage.In this paper,we study the image segmentation based on unit-linking pulse coupled neural network. Here we apply histogram to the threshold selection of neuron pulse capture process, propose a new evaluation criteria, and extend the unit pulse coupled neural network to three-dimensional space. The main work and contribution of the paper contains the following aspects:1.We introduce a histogram threshold-based image segmentation method, combining the histogram and unit-linking pulse coupled neural network. The method makes full use of the prior gray distribution of original image to solve the optimal threshold selecting problems. Meanwhile, compared to the traditional multi-threshold segmentation method, it makes full use of the spatial location of the target information. When there exists small gap between the target and background information, it could split the target and improve the accuracy of image segmentation. 2.We propose a new image segmentation standard called the edged product mutual information criterion. This criterion does not require human intervention, and rely on the computer’s evaluation. How to determine the optimal separation time of the background and the target is a key factor in image segmentation. Among all existing evaluation methods, mutual information is widely used in image segmentation because it could achieve good results. However,it caused the poor effect in the condition that there exists blurred image background and objectives. Besides, the mutual information would involve a large number of floating point operations to influence the effectiveness. Our method introduces edge information and simplify the logarithm of the mutual information. Experiments show that our method could retain the image edge details and improve the efficiency greatly.3. We apply the proposed unit-linking pulse coupled neural network model based on histogram and the edged product mutual information criterion to realize the whole segmentation in three-dimensional space.3D whole segmentation method takes full advantage of the location information. Segmenting image directly in three-dimension space would take the characteristics of the surrounding pixels into account, and its result would be related to the space adjacent to other pixels. Compared to3D Otsu method, our method could improve the segmentation accuracy. Besides, it is much easier in hardware implementation and better in accuracy than traditional3D pulse coupled neural network.