Research On Image Classification And Fusion Based On Machine Learning Techniques

Images have always played an essential role in human's life since vision is most likely human beings' important sense.As a result,the area of image processing has numerous fields such as remote sensing,military,medical,etc.These days,generating a huge number of images is very easy for everyone to capture anywhere with digital technologies.The conventional techniques of image processing have to deal with more complicated issues of images and also facing to adapt according to the human vision.Machine learning techniques have an advanced and significant component in image processing development.Recently image processing and machine learning have obtained great attention due to benchmarks and image datasets.Machine learning in image processing has Innovative integration that has a lot of advantages,which has provided a better understanding of complicated images.In this dissertation,three different projects are performed in image processing,such as pan-sharpening of multi-spectral(MS)and panchromatic(PAN)images,multi-focus image fusion and hyperspectral image classification with novel machine learning techniques.This study proposes the image enhancement before pan-sharpening;that is,the image enhancement methods are applied as a pre-processing step.The image enhancement methods are proposed in two domains,i.e.,"same-domain and cross-domain".In the same-domain methods,the image enhancement techniques(such as Laplacian,Unsharp)are simply applied to multispectral and panchromatic images to sharpen both images in the spatial domain.While in cross-domain,a novel hybrid combination of Laplacian Filter(LF)and Discrete Fourier Transformation(DFT)image sharpening technique is introduced.After image enhancement,the powerful Matting Model(MM)pan-sharpening technique is used to fuse both the enhanced images and produce a resultant image with high spatial and spectral resolutions.The experimentation results of the proposed approach outperform the others as compared to the benchmarks techniques over two image sets.The results are evaluated,considering both Qualitative and Quantitative evaluation metrics.The Multi-focus image is "limited Depth-of-Field(DOF)of an imaging system causes blur images when the sample is wider than the DOF of the optical system".Color multi-focused image fusion allows merging two multi-focus images and generates a composite image integrating complementary information to understand the entire scene better.This study introduced a new two-level approach for color image fusion.In the first level,the multi-focus image is enhanced by the Laplacian Filter(LF)technique with a Discrete Fourier Transform(DFT).In the second level,the enhanced images are further processed by Stationary Wavelet Transform(SWT)and construct a more informative fused image.The experimental results on the color multi-focused images showed the capabilities,and improved results of the novel approach are compared with the traditional SWT method.The output image is evaluated using both a qualitative and a quantitative approach.The eight quantitative metrics are applied to evaluate the performance of the novel technique.The article also proposed a new type of qualitative performance metric named as "Qualitative Error Image(QEI)" to evaluate the proposed method and assess future evaluation.Nearest regularized subspace(NRS)has been recently proposed for hyperspectral image(HSI)classification.The NRS outperforms both collaborative representation classification and sparse presentation-based techniques because the NRS makes use of the distance-weighted Tikhonov regularization to ensure appropriate representation from similar samples within-class.However,typical NRS only considers Euclidean distance,which may be suboptimal to resolve the problem of sensitivity in the absolute magnitude of a spectrum.An NRS-Manhattan distance(MD)strategy is proposed for HSI classification.The proposed distance metric controls over magnitude change and emphasizes the shape of the spectrum.Furthermore,the MD metric uses the complete information of the spectral bands in full dimensionality of the HSI pixels,which makes NRS-MD a more efficient pixel-wise classifier.Validations are done with several hyperspectral data,i.e.,Indian Pines,Botswana,Salinas,and Houston.Results show that the novel NRS-MD is better than other benchmarks techniques.