Research On Image Segmentation And Fusion Based On Pulse Coupled Neural Network

The pulse coupled neural network(PCNN)is a network model for simulating the characteristics of mammalian optic nerve cells proposed in the early 1990 s.Due to its own characteristics of biological vision such as variable threshold,capture,automatic wave and integrated space-time,it is very suitable for application in image processing.An intensive study has been conducted on PCNN principle and model in this paper.Aimed at problems both in PCNN image segmentation and PCNN image fusion,such as the too many parameters,difficulties in setting and the unspecific and fragmentary segmentation evaluation for the image segmentation,as well as the inaccurate division for clear region,the poor anti-noise performance of regional definition characteristics in tradition and the low-quality fusion results for image fusion,PCNN was combined with the particle swarm optimization(PSO)and nonsubsampled countourlet transform(NSCT)algorithm so as to put forward the new image segmentation method and image fusion algorithm.Multi-parameter setting and single segmentation evaluation criterion are the problems in image segmentation based on PCNN.Through combining particle swarm optimization(PSO)with comprehensive evaluation criterion,this paper presents an automatic image segmentation algorithm based on PCNN.The improved PCNN model with monotonically increasing threshold search strategy was utilized in this algorithm.The comprehensive evaluation criterion(CEC)obtained by cross-entropy parameter,edge matching degree and noise control degree was proposed as the fitness of particles in PSO,then the parameters of PCNN such as the target time constant,the connection coefficient and the iteration times n are acquired adaptively by updating fitness value of particles.Using these acquired optimum parameters,the image was segmented by the improved PCNN.For different types of images,experimental results show that our algorithm can segment image completely and accurately under PCNN operating efficiency,moreover texture details are retained.Compared with other experimental methods,our segmented results are superior to other algorithm in CEC.In addition,the general comprehensive indicators of our segmented results are also optimal.Thus we can see the objective evaluations are consistent with the visual subjective evaluations,and the algorithm has high robustness.In image fusion,this paper combined with nonsubsampled countourlet transform(NSCT)and dual-channel PCNN(DCPCNN),proposed an image fusion method.The model was modified by DCPCNN with using the internal activity item of fraction and segment-based dynamic threshold.Multiscale transformation of fused source images was performed using NSCT.Low-frequency fusion was conducted based on DCPCNN.The normalized low-frequency-band coefficient and sum of fractal feature area(SFFA)of source images were input as external stimulation and connection coefficient,respectively.Time matrix was output,fusion template image FTI was confirmed,and low-frequency-subband coefficient was obtained.High-frequency fusion was carried out using Sum Modified-laplacian(SML)for the obtaining of high-frequency-subband coefficient.In multi-focus image fusion,a comparative study was carried out.The results of experiment in low-frequency coefficient fusion template showed that the proposed DCPCNN algorithm with internal activity item of fraction and segment-based dynamic thresholds in this paper could accurately select low-frequency fusion coefficients from the pixel area with gentle gray change,with high stability.In multi-focus image fusion,this method was the optimal in both experiment groups with standard reference fusion images and without reference standard reference fusion images,showing coincident subject visual sense and data analysis,high stability and ideal fusion result.