Research Of Feature-level Image Fusion Algorithm Based On Fuzzy Neural Network

With the development of sensor manufacturing technology and the explosive growth of information,information fusion technology has become an important research branch of information processing.Since the information from different sources are varietal,diverse and complementary,information fusion aims at fusing multiple sources of information into ones to get more abundant,correct and complete information than single source.Image is an important manifestation of information and image fusion has become a research hot spot in information fusion.After analyzing and studying some classical algorithms of image fusion,an improved algorithm is proposed.The main contents of this paper include the following aspects:(1)This paper presented a feature-level multi-focus image fusion framework based on fuzzy neural network.First,the source image is segmented into some blocks with same size.Then the redundancy classification of the image blocks is made.Next,the clarity characteristics in a non-redundant image block are extracted and the classification is made by the fuzzy neural network.Finally,the image blocks are fused according to the classify results.Experiments show that the proposed framework is reasonable,the processing speed is fast,and the fusion results are better than classical pixel level fusion algorithms.(2)The classification of the image blocks is made by using the fuzzy neural network combined with the SVM.Firstly,the membership function parameters of the fuzzy neural network antecedent are got by FCM clustering.Then the output equation parameters of the fuzzy neural network consequent are got by SVM training.Thus the combination of SVM and fuzzy neural network is achieved organically.Experimental results show that this method reduces the number of fuzzy neural network rules and the training time of parameters and improves the classification accuracy.(3)A method to optimize the parameters of fuzzy neural network by using modified PSO is proposed.The fitness function is designed based on the average absolute error of fuzzy neural network output and standard output and the adaptive inertia factor is designed according to the local and global fitness.Thus the global optimization and local smooth convergence are achieved.Experiments show that the improved PSO algorithm has faster convergence and higher classification accuracy than the BP algorithm and has more stable and smooth optimization process than linear inertia factor PSO algorithm with higher classification accuracy.