Signal processing algorithms for enhanced image fusion performance and assessment

摘要

The dissertation presents several signal processing algorithms for image fusion in noisy multimodaludconditions. It introduces a novel image fusion method which performs well for imageudsets heavily corrupted by noise. As opposed to current image fusion schemes, the method hasudno requirements for a priori knowledge of the noise component. The image is decomposed withudChebyshev polynomials (CP) being used as basis functions to perform fusion at feature level. Theudproperties of CP, namely fast convergence and smooth approximation, renders it ideal for heuristicudand indiscriminate denoising fusion tasks. Quantitative evaluation using objective fusion assessmentudmethods show favourable performance of the proposed scheme compared to previous effortsudon image fusion, notably in heavily corrupted images.udThe approach is further improved by incorporating the advantages of CP with a state-of-the-artudfusion technique named independent component analysis (ICA), for joint-fusion processingudbased on region saliency. Whilst CP fusion is robust under severe noise conditions, it is prone toudeliminating high frequency information of the images involved, thereby limiting image sharpness.udFusion using ICA, on the other hand, performs well in transferring edges and other salient featuresudof the input images into the composite output. The combination of both methods, coupled withudseveral mathematical morphological operations in an algorithm fusion framework, is considered audviable solution. Again, according to the quantitative metrics the results of our proposed approachudare very encouraging as far as joint fusion and denoising are concerned.udAnother focus of this dissertation is on a novel metric for image fusion evaluation that is basedudon texture. The conservation of background textural details is considered important in many fusionudapplications as they help define the image depth and structure, which may prove crucial inudmany surveillance and remote sensing applications. Our work aims to evaluate the performance of image fusion algorithms based on their ability to retain textural details from the fusion process.udThis is done by utilising the gray-level co-occurrence matrix (GLCM) model to extract second-orderudstatistical features for the derivation of an image textural measure, which is then used toudreplace the edge-based calculations in an objective-based fusion metric. Performance evaluationudon established fusion methods verifies that the proposed metric is viable, especially for multimodaludscenarios.