Study On Some Key Techniques In Image Fusion

Image fusion is an important research task in pattern recognition and image processing. Due to the pattern’s diversity in the real world, it is impossible to find a common method suitable to all scenes, which brings a lot of challenges for the image fusion task.In this study, several image fusion methods are addressed based on different techniques including dual-channel PCNN,sparse representation,fuzzy system,hidden markov model. Moreover, these new methods are evaluated in terms of both visual quality and quantitative evaluation metrics. In general, the work addressted in this study includes the following several aspects.(1) A novel method for self-adaptive dual-channel pulse coupled neural networks(DC-PCNN) based on particle swarm optimization(PSO) evolutionary learning is proposed in order to overcome the difficulty of parameters selection of DC-PCNN. In this study an evolutionary learning algorithm and a new optimization criterion are proposed to optimize the parameters of PCNN for image fusion. In contrast with classical DC-PCNN method that needs to try different parameters settings manually, the proposed method can find the optimal parameters adaptively.(2) Compressed sensing(CS) has attracted a lot of attention in recent years, and it remains a hot research topic in pattern recognition and image processing. In this paper, a novel method by integrating compressed sensing with pulse coupled neural networks(PCNN) is put forward for the purpose of overcoming the drawback of noise sensitivity and poor time efficiency of the classical PCNN. The proposed method not only has good ability to overcome the noise, but also performs the denoising and image fusion simultaneously, whereas denoising and fusion processes are carried out separately for many conventional image fusion approaches and this would result in information inconsistency. Nevertheless, by integrating the merits of CS and PCNN, the proposed method can greatly improve the image fusion efficiency and reduce the computation time to some extent.(3) For multi-channel image fusion, the corresponding pulse coupled neural network image fusion methods are still very scarce. When the tradition dual channel pulse coupled neural network image fusion methods are used for the multi-channel image fusion, the results are easily influenced by different fusion sequence. To overcome this difficulty, a novel multi-channel pulse coupled neural network(M-PCNN) method is proposed. Meanwhile, a novel calculation method about linking-weight based on gray energy is also proposed for multi-channel image fusion in order to overcome the lack of linking-weight computation method in M-PCNN, which can realize the cooperation of different channels, reduce the information loss, avoid the Laplacian operator calculation repeatedly and have the better computation.(4) In order to overcome the shortcoming caused by the equal weighting on LL low-low Frequency(LL) components, low-high Frequency(LH) components of low resolution images and high Frequency(H)components in the sparse representation image fusion model, a novel Shannon entropy multi-view weighting based sparse representation image fusion method is proposed. The proposed method can assign the different weights to LL, LH and H components, and adaptively enhance the influences of the important components. Thus, the image fusion effect can be improved effectively.(5) A novel image fusion framework based on supervised intelligent learning is proposed in order to overcome the lack that supervised learning mechanism and priori knowledge cannot efficiently be used in the image fusion procedure. In this study, the images database for supervised learning is effectively constructed according to priori knowledge(i.e. high-grade fusion image) before image fusion and used for the model parameters training of the available classical supervised learning models, and then the trained model is used to guide the new fusion task.(6) A novel image fusion method based on supervised intelligent learning in fuzzy system is proposed in order to overcome the difficulty in the use of priori knowledge in image fusion. In this study, the images database for supervised learning is first constructed,and then the model parameters trained with the available training datasets are used for the takagi sugeno kang(TSK) fuzzy system model. Meanwhile, some advantages are displayed in the fusion image quality and adaptation.(7) A novel image fusion framework based on HMM is proposed in order to overcome the lack that the connected relationship in adjacent points cannot efficiently be used in the image fusion procedure. In this study, the probability is constructed according to the mean and average gradient. Fusion images are built by back propagation(BP) optimization. Differentiated from the classical method that needs to isolate the fusion process in more near points, the proposed method can effectively avoid the difficulty and improve the fusion performance.