Research On PCNN Model For Infrared Human Segmentation Under Complex Environment

In recent years,infrared human detection technique is widely used in intelligent security and surveillance,traffic navigation and recognition,medical image analysis,military target detection and other applications,because of its excellent characteristics of strong disguise,all-weather capabilities and high sensitivity with thermal targets.In this research area,Pulse Coupled Neural Network(PCNN)has palyed an increasingly significant role in enhancing the performance of the infrared image segmentation,and embraces a single-layer network with simple architecture,free training mechanism,pulse synchronization and acquisition of the neurons.PCNN is based on the simulation of the photosensitivity mechanism of the mammal visual cortex neurons in order to segment the human targets from infrared images effectively,and is different from other traditional image segmentation methods.In this network,each neuron can fuse external input signals,iterative output signals in time scale and neighborhood output signals from other neurons in space scale by the nonlinear coupled mode,and can synchronously fire based on the same properties.Accordingly,the PCNN model can segment infrared human targets in a complex background by the pulse generation.At present,many domestic and foreign researchers make the effort to improve and optimize the internal structure of PCNN for image segmentation applications.Additionally,several improved models are emerged with optimization parameters,simplified internal mechanisms and different application requirements.However,in the application of infrared human segmentation in real scene,the current PCNN mode still have so many shortcomings including the blurred edges of segmentation results,the serious loss of human details,the over-sensitivity of infrared noise,the complex structure,and extraneous parameter setting mechanism,which has become the key problems and constrained the further development of PCNN in this research area.Therefore,the related research work in this thesis has been carried out in order to address the above key problems.And the main contributions are listed as follows:(1)The operation mechanism of the current PCNN models is analyzed in the process of the infrared image segmentation,and a set of simplified PCNN model frameworks for infrared human target segmentation is proposed.In these model frameworks,firstly,the appropriate values of the linking strength in the modulation field can be set adaptively based on the infrared feature information of human targets.Moreover,the gray-level distribution information of the iterative segmentation region is employed to build the dynamic threshold to simplify the pulse generation mode.Lastly,on the basis of thermal imaging characteristics of infrared human,the relative change of the average gray value of the output region is detected in the iterative process of the PCNN model,so as to terminate this segmentation process automatically.(2)Existing PCNN models commonly have several problems including the blurred edges of targets,the serious loss of human details and the inaccuracy of segmentation.Curvature Grey-level Gradient Tensor(CGGT)is brought forward to improve the PCNN model in order to address these problems,and can express the edges and details of human target more exactly in segmentation.Experiments on 200 real infrared images,demonstrate that this method is superior over the other classic segmentation methods in both the subjective visual performance and the objective indicators including the average misclassification error of 0.59% and the average f-measure score of 0.89,and is effectively adapt to the relatively poor segmentation conditions of the more complex background interferences,the smaller targets,the blurred regional boundary,and the crosses between targets.(3)Currently,traditional PCNN models are also facing the challenge of the poor adaptability of infrared noise.Based on the generation mechanisms and features of the infrared noise in the thermal imaging process,an improved PCNN model is presented to integrate the noise suppression module which is designed as the weight matrix of the feeding input field by the anisotropic Gaussian kernels(ANGKs)and the visual salience information which is defined as the normalized spectral residual saliency by the linking strength to enhance the contour expression of infrared human targets,aiming at alleviating the problem of the over-sensitivity of infrared noise.Experiments on 100 infrared noise images with different Peak Signal to Noise Ratio(PSNR)levels,show that the superiority in quantitative evaluations of the average misclassification error(ME),the average variation of information(VI)and the average probabilistic rand index(PRI)for the proposed PCNN model,compared with other classic improved PCNN models.(4)In order to deal with the parameter estimation problem of PCNN in the infrared human segmentation,the improved cuckoo search(CS)algorithm is devised as a swarm intelligent optimization method.This method introduces the sophisticated local search based on the Kentchaotic map and the guidance mechanism of the k dimensional tree to the standard CS algorithm,and overpass the drawbacks of the weak ability of the local search and the search blindness of the standard model.Compared to other PCNN segmentation models with existing swarm intelligent parameter optimizations,the proposed algorithm gets the superiority in the visual segmentation effect and the computing efficiency.And experiments on 100 infrared images,demonstrate that the average misclassification error of the proposed algorithm is at least 1% lower than other methods,and the average executive time of this algorithm is at least 2 seconds faster than other methods.In addition,two other improved PCNN models proposed in this paper also are introduced in this experiment including the improved PCNN model combined with CGGT(CGGT-PCNN)and the improved PCNN model for infrared noise suppression(Denosing-PCNN).Experimental results show that the PCNN model with improved cuckoo search has no significant differences with CGGT-PCNN and Denosing-PCNN in the application of infrared human image segmentation with high signal to noise ratio,but is not adapted to infrared human segmentation under strong noise background.This algorithm is relatively time-consuming because of the swarm intelligence optimization process.