https://www.um.edu.mt/library/oar/handle/123456789/22542 Sat, 06 Jun 2026 13:33:14 GMT 2026-06-06T13:33:14Z https://www.um.edu.mt/library/oar/handle/123456789/38983 Title: Hand drawn sketch drawings to vector graphics Abstract: Computer Numerical Control (CNC) machines and 3D printers are becoming more accessible. This allows artists and consumers to create physical objects from their drawn ideas. However, many artists are more familiar with the traditional method of sketching which uses raster format. For these kind of devices, a special instruction set based on vector notations is used which requires specialised software. One problem is that learning a new interface or technology for 3D modelling is not trivial and some might give up before trying. A bridge that easily connects these two worlds would be beneficial to both. However, even though the interpretation of sketches appears to be trivial for humans, so cannot be said for machines. This holds particularly true when artists use artistic cues such as shadows to represent depth. Additionally, hand drawn sketches are intrinsically imperfect and might contain curves making the gap between raster to vector hard to reduce. In this research, a method that automatically converts hand-drawn sketches in presence of shadows and curves is presented. The following study is divided into two section. The first section deals with junction localisation and identification to ensure that the topological fidelity of the drawing is retained. When compared to current state of the art, the results obtained shows an improvement of 61% when the proposed methodology was evaulated for junction spatial localisation using Salient Point Error over the same dataset. Even though junction type identification was not used during the proposed vectorisation pipeline, a number of methods were described and evaluated for junction type classification. Classification was performed using three methodologies and the best classification results obtained an F-score of 0.95. The second section dealt with contour extraction to remove shadows and other artefacts from the drawings. Each pixel was assigned an orientation based on the direction of the surrounding pixels and used the result was used to identify the path between two connected junctions. Unconnected lines and recovery of missed junctions were also considered. An average F-measure score of 0.992 was obtained over the whole dataset when the ground truths and the vectorised images were compared using a contour evaluation protocol. A dataset of 17 images was used to cover drawings created in sketching software and on paper using straight and curved lines. The reconstruction was performed using either lines, arcs or splines as deemed the most adequate. We showed that our method performed better in junction and contour detection and the results obtained were consistent throughout the whole dataset including straight and curved lines drawings. Description: M.SC.ARTIFICIAL INTELLIGENCE Sun, 01 Jan 2017 00:00:00 GMT https://www.um.edu.mt/library/oar/handle/123456789/38983 2017-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/38859 Title: Semi-automatic segmentation of human anatomical imagery Abstract: Manual segmentation of anatomical imagery is a challenging and laborious task which this dissertation attempts to alleviate. We present a semi-automatic segmentation system which operates on a new data set of photographic human anatomical imagery. A morphological tree-based segmentation method was utilised in order to reach this aim. We placed a particular focus on elongated structures in order to demonstrate the e ectiveness of the algorithms. The resultant outputs were presented to academics in the anatomical sciences for evaluation. Qualitative and quantitative results which were collected throughout the course of the experimentation phase indicate that the system was successful in producing meaningful labelled segmentation outputs with particularly good performance on elongation, which were commended by the experts. We believe that these results provide a good initialisation step for more re ned labelled images which can be used in a number of di erent professional and educational tools. Furthermore, the outcome of this dissertation demonstrates that a technical window exists in this area, and a foundation for further research has been created in this work. Description: M.SC.ARTIFICIAL INTELLIGENCE Mon, 01 Jan 2018 00:00:00 GMT https://www.um.edu.mt/library/oar/handle/123456789/38859 2018-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/36180 Title: Multiframe blind image deconvolution with space and time variant PSFs Abstract: The study of images in scientific fields such as remote sensing, medical imaging, and astronomy is important not only because pictures mimic one of the main human sensory modalities, but also because, in the case of signals outside the visible wavelengths, they allow for the visualization of artefacts beyond the sensitive range of the human eye. However, accurate information can only be extracted if the data is free from noise, and perhaps more importantly – blur. An improper focal-length setting, insufficient light, or camera motion, can introduce additional filtering during the image capture process. Medical image that require long exposure time can, for instance, lose sharpness due to patient movement. In radio astronomy, interferometers with unprecedented sensitivity, large fields of view, and narrow angular resolutions, are being built. However, the finite set of baselines still makes the accurate rendering of sharp radio images difficult. Similarly, the larger the aperture of an optical telescope, the stronger the aberrations introduced by the lens. Apart from hardware limitations, biases also arise from phenomena beyond human control such as, for instance, turbulence in the atmospheric and the ionospheric columns that change on millisecond scales. Induced distortions in the images are defined by the Point Spread Function (PSF) that models how rays from a point source trace in the instrument. The deconvolution process attempts to undo these adverse effects and recover the actual intensity values from the measured signal. The blur filter varies in time depending on the astronomical ‘seeing’ conditions as well as on the wavelengths being recorded. While in certain cases expensive hardware such as adaptive optics can help to elevate the problem, a software-based approach is more common due to its wide-ranging adaptability and applicability. This study proposes techniques that render a sharp version of a scene using multiple blurred frames captured over the same area. Each image is assumed to have a different, but unknown PSF. Kolmogorov and Moffat kernel that model the atmospheric and ionospheric effects as well as the filtering due to the lens or interferometry, are used. Many current methods adopt a non-blind approach (in the sense that the PSF is known) or estimate the blur filter form edge regions. This restricts their applicability to specific types of images. For instance, cosmological frames that encode faint point sources or extended object with a complex morphology, cannot be processed. Moreover, the majority of existing multiframe algorithms work in batch mode and assume the availability of all frames in the pre-processing stage. The objective of this study is to propose methods for Multiframe and Blind Image Deconvolution (MBID) – with or without noise. The penultimate goal is to reverse the effects of convolution without knowledge of the PSF. Complementary information from multiple images is combined to estimate a single sharp version of the scene. The deconvolution problem is reduced to a series of iterations in which the blur kernel is initially estimated and subsequently used to deconvolve the input frame. The result is in turn used to update the latent image. Reversing the spreading introduced by the convolution operator is an ill-posed problem. Finding a solution for x from the measurement set y in the undermined linear system y = Ax + ɛ is challenging. Infinitely many correct answers remain possible even if the PSF (A), and noise level (ɛ) are known. The frameworks we present make use of novel theories developed in the field of Compressed Sensing. Sparse magnitude coefficients are obtained by converging towards the vector with the least quadratic error that minimizes the l2 norm, or by searching for the sparses solution through the minimization of the l1 norm. Various methods including genetic programming and subtractive minimization in different domain representations, are considered for the recovery of the non-sparse but constrained phase information. An evolutionary based strategy that allows the performance to be tested at various stages of the deblurring process is designed and followed. Quantification of the recovery accuracy is made possible through the computation of various error metrics. A comparison with the images recovered by the current state-of-the art methods when the same set of blurred frames are processed, is also put forward. This analysis also takes into account the robustness to various noise levels. Encouraging results that are comparable and, at times, even better than those achieved by the existing approaches, are obtained. Description: PH.D.ARTIFICIAL INTELLIGENCE Sun, 01 Jan 2017 00:00:00 GMT https://www.um.edu.mt/library/oar/handle/123456789/36180 2017-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/30050 Title: Enhancing android malware sandboxes with anti-evasion code patching Abstract: Sophisticated Android malware families often implement techniques aimed at avoiding detection. Split personality malware for example, behaves benignly when it detects that it is running on an analysis environment such as a malware sandbox, and maliciously when running on a real user's device. These kind of techniques are problematic for malware analysts, often rendering them unable to detect or understand the malicious behaviour. This is where sandbox hardening comes into play. In this work, we exploit sandbox detection heuristic prediction to proactively generate bytecode patches, in order to disable the malware's ability to detect a malware sandbox. Through the development of AndroNeo, we demonstrate the feasibility of this approach by showing that the heuristic prediction basis is a solid starting point to build upon, and demonstrating that when heuristic prediction is followed by bytecode patch generation, split personality can be defeated. The AndroNeo prototype implements checks at the Java level for API method calls that can distinguish real devices from emulators. The robustness of AndroNeo was demonstrated by showing its ability to identify and patch evasion heuristics within packed code. The relevance of packed malware was confirmed by demonstrating the prevalence of packers in modern day malware samples. Description: M.SC.COMP.SCI.&ARTIFICIAL INTELLIGENCE Sun, 01 Jan 2017 00:00:00 GMT https://www.um.edu.mt/library/oar/handle/123456789/30050 2017-01-01T00:00:00Z