Pulse Coupled Neural Networks, Image Fusion Of Multi-scale

Image fusion aims to get a synthesis image satisfying human visual perception. Recently, due to such perfect properties as Multi-Scale and Time-Frequency localization, the wavelet transform has been widely used in the field of image fusion. But for the two-dimension structure of image data, the wavelet transform cannot effectively represent the important information in the image such as 'line', 'curve', 'edge' and so on. Multi-Scale geometric analysis method of nonsubsampled contourlet transform (NSCT) is considered a true representation of image. It not only inherits the traditional wavelet transform properties of Multi-Scale and Time-Frequency localization but also has Multi-direction and anisotropy. Furthermore, it has the shift invariance property which overcomes the Ring Artifacts caused by sampling operation. Pulse-Coupled Neural Networks (PCNN) is a visual cortex-inspired neural network and has its mechanism of information processing and signal conduct. So it has been introduced in image fusion field. In a word, the method of image fusion which based on Multi-Scale decomposition and improved PCNN has a larger application value and be focused on in this paper. So contents and conclusions of this paper are as follows:Firstly, theories of Wavelet Transform, Contourlet Transform (CT) and NSCT, and principles of how to structure and design the wavelet filters are recounted, and those characteristics and physical meaning in the specific application be summarized. Then, the PCNN's theory together with parameters setting and the way of working are analyzed in detail. Finally, some conclusions associated with the purpose of image fusion and its properties are obtained from all kinds of experiments.Secondly, to overcome the limitation of the image fusion method based on Multi-Scale wavelet transform, an advanced Local-Scharr operation method in wavelet transform domain is proposed. This Scharr operation has better approximation property and the best rotation invariance than other schemes. Experiments have shown that this method has consistent advantages and robustness.Thirdly, as the traditional method based on Multi-Scale and PCNN without considering its neighborhood information to motivate PCNN neurons, experiment results have low contrast in edges and Boundary Effect owing to either-or method to select the best coefficient which has large firing times in PCNN model. So a novel method based on adaptive Local Spatial Frequency (LSF) in NSCT domain and PCNN (NSCTLSFPCNN) is proposed because of LSF standing the detail information of the image and adaptive fusion rules eliminating the Boundary Effect effectively, experiments show its effectiveness.Fourthly, under the image fusion framework based on Multiscale and PCNN, the optimal method (NSCTLSFPCNN) together with its optimal five key parameters and some setting principles has found by experiments respectively.