|Institution:||University of New South Wales|
|Department:||Engineering & Information Technology|
|Keywords:||Biometric Image; Image Denoising; Biomedical Image|
|Full text PDF:||http://handle.unsw.edu.au/1959.4/54348|
Image denoising techniques are important to cope with different types of noise in biomedical and biometric images. Not only random noise but also general irrelevant components in image spatial domains have been treated as undesired information. To reduce or eliminate the adverse effects of random noise and irrelevant components, this thesis works on following four real-life problems encountered in biomedical and biometric image applications. Firstly, microscopy images are often degraded by random noise from readout procedures and image data acquisition systems, devices or equipment. This thesis proposes an integration of trend surface mapping, Q-Q plot, bootstrapping, and Gaussian spatial kernel for removing Gaussian-like noise in microscopy images. Furthermore, the proposed approach can be extended to handle Poisson noise. Experiments on synthetic and real noise datasets demonstrate the advantages of the proposed method. Secondly, medical image classification is challenged by concurrent occurrence of image rotation change, scale variation and noise corruption. This thesis introduces two image features, named particle potential motion entropy histogram (PPMEH) and its updated version PPMEH-FT, incorporated with discrete Fourier transform (DFT), to deal with the multiple effects of rotation, scaling and noise for classifying medical images. The experiments on computed tomography (CT) and magnetic resonance imaging (MRI) datasets show that the proposed image features outperform state-of-the-art methods. Thirdly, latent fingerprints are usually small-sized, blurred, and overlapped with irrelevant image components. Segmentation of latent fingerprint is very challenging under complex and poor quality image conditions. Furthermore, subsequent latent fingerprint matching is another difficulty. This thesis developed a fully automated latent matcher embedded with a robust latent segmentation module and experiments with a latent fingerprint database demonstrate superiority of the proposed multi-module latent matcher. This thesis targets three real-life problems in biomedical and biometric image applications. Performance evaluation and comparison with current state-of-the-art approaches validate that the proposed techniques are effective solutions to such problems.