Research On Multi-Model And Multi-Focus Image Fusion Based On Multi-Scale Decomposition And Feature Extraction Method

Image fusion(IF)is the process of amalgamating multiple input images to obtain a final fused image that extracts all substantial features with better contrast without introducing artifacts and noise.With the recent advancement of multi-scale decomposition(MSD)and deep learning(DL)methods,the IF has acquired remarkable success due to its diverse applications such as computer vision,object detection,surveillance,medical imaging,and so on.Various attempts have been made to produce a final fused image with better results;however,the fused image still lacks in extracting useful information due to variations in the background contrast,environmental changes,blurring effect,improper fusion strategy,uneven illumination and presence of noise and artifacts.In this dissertation,four new image fusion methods have been proposed for the general image fusion(IF)framework,multi-focus IF,infrared(IR)and visible(VI)IF,and medical IF to address the aforementioned issues.The first part of this research work deals with designing a hybrid image fusion method for a general image fusion framework which is suitable for all categories of IF,such as multi-focus IF,IR-VI image fusion,and CT-MRI image fusion.The fundamental idea of this hybrid method is to investigate the properties and characteristics of IF categories(Multi-focus IF,IR-VI image fusion and CT-MRI image fusion)so that we can examine the suitable domain for each category more specifically.This method uses a combined median-average filter-based hybrid discrete stationary wavelet transform(DSWT)and principal component analysis(PCA)for a fusion of images.This method overcomes the trade-off of Heisenberg’s uncertainty principle by improving accuracy in both domains(spatial and spectral)by producing output image with sufficient energy information and enriched detailed features with negligible noise.The second part of this research deals with multi-focus IF,which uses amalgamated histogram equalization and fast gray-level grouping(HE-FGLV)along with non-subsampled contourlet transform(NSCT)and grey principal component analysis(G-PCA).The HE-FGLG automatically improves contrast with its fast computation;the NSCT decomposes the images in low-frequency(LF)and high-frequency(HF)sub-bands by retaining the true geometrical structure of an image,such as edges and contours.Then,local energy and mean gradient fusion strategies are used to restore energy information with enriched detailed features.Finally,GPCA based multi-channel color to single-channel gray conversion technique is implemented that preserve better contrast in gray-scale while it also helps in dimension reduction.The third part of this research deals with IR and VI image fusion,which uses an adaptive fuzzy-based preprocessing enhancement method,anisotropic diffusion(AD)filtering scheme,VGG-19 network through the guidance of transfer learning,and multiple fusion strategies.An adaptive fuzzy-based preprocessing enhancement method is employed that automatically enhances the contrast of images with adaptive parameter calculation.The enhanced images are then decomposed into base and detail layers by anisotropic diffusion(AD)based edgepreserving filters that remove noise while smoothing the edges.The detailed parts are fed to four convolutional layers of the VGG-19 network through transfer learning to extract feature maps by multiple fusion strategies to get the final fused image enriched with detailed features.In contrast,the base parts are fused by the PCA method to preserve the energy information.The fourth part of this work deals with computed tomography(CT)and magnetic resonance imaging(MRI)medical image fusion,which uses a convolutional neural network(CNN),local shift-invariant shearlet transform(LSIST),and preprocessing method employing bottom-hattop-hat(BHTH)along with grey-PCA.Firstly,the preprocessing methods address the uneven illumination and poor contrast issues.In contrast,the LSIST method decomposes the images into the LF and HF sub-bands,efficiently restoring all substantial features in various scales and directions.Afterward,the HF images are fused by the CNN method that captures smooth edges,textures and contours without introducing artifacts.In contrast,the LF images are fused by the local energy fusion rule to preserve energy information.To summarize,this work concludes that four distinct effective IF methods have been proposed;one for the general image fusion framework and three for distinct categories that are multi-focus,IR and VI,and medical image fusion.The proposed fusion methods produce a final fused image that extracts all important information from individual source images(SI)with very negligible noise.The experimental results reveal that proposed methods achieve state-of-the-art(SOTA)performance than existing fusion methods in subjective evaluation through the visual experience of experts and objective evaluation parameters.