Search

Content-based image classification and retrieval are the automatic processes of taking an unseen image input and extracting its features representing the input image. Then, for the classification task, this mathematically measured input is categorized according to established criteria in the server and consequently shows the output as a result. On the other hand, for the retrieval task, the extracted features of an unseen query image are sent to the server to search for the most visually similar images to a given image and retrieve these images as a result. Despite image features could be represented by classical features, artificial intelligence-based features, Convolutional Neural Networks (CNN) to be precise, have become powerful tools in the field. Nonetheless, the high dimensional CNN features have been a challenge in particular for applications on mobile or Internet of Things devices. Therefore, in this thesis, several fast embeddings are explored and proposed to overcome the constraints of low memory, bandwidth, and power. Furthermore, the first hostel image database is created with three datasets, hostel image dataset containing 13,908 interior and exterior images of hostels across the world, and Hostels-900 dataset and Hostels-2K dataset containing 972 images and 2,380 images, respectively, of 20 London hostel buildings. The results demonstrate that the proposed fast embeddings such as the application of GHM-Rand operator, GHM-Fix operator, and binary feature vectors are able to outperform or give competitive results to those state-of-the-art methods with a lot less computational resource. Additionally, the findings from a ten-year literature review of CBIR study in the tourism industry could picturize the relevant research activities in the past decade which are not only beneficial to the hostel industry or tourism sector but also to the computer science and engineering research communities for the potential real-life applications of the existing and developing technologies in the field.

With the rapid development in video coding technologies in the last decade, high-resolution video delivery suffers from packet loss due to unreliable transmission channels (time-varying characteristics). The error Resilience approaches at channel coding level are less efficient to implement in real time video transmission as the encoded video samples are in variable code length. Therefore, error resilience in video coding standard plays a vital role to reduce the effect of error propagation and improve the perceived visual quality. The main work in this thesis is to develop an efficient error resilience mechanism for H.265|HEVC video coding standard to reduce the effects of error propagation in error-prone conditions. In this thesis, two error resilience algorithms are proposed. The first one is Adaptive Slice Encoding (ASE) error resilience algorithm. The concept of this algorithm is to extract and protect the most active slices in the coded bitstream based on the adaptive search window. This algorithm can be applied in low delay video transmission with and without using a feedback channel. It is also designed to be compatible with reference coding software manual (HM16) for H.265|HEVC coding standard. The second proposed algorithm is a joint encoder-decoder error resilience called Error resilience based on Supplemental Enhancement Information (ERSEI) algorithm. A feedback message status is used from the decoder to notify the encoder to start encoding clean random-access picture adaptively based on the decoded picture hash message status from the decoder. At the same time, the decoder will be notified to start the error concealment process whilst waiting to receive correct video data. A recovery point message from the decoder feedback channel is used to update the encoder with error messages. In this thesis, extensive experimental work, evaluation, and comparison with state-of-the-art related algorithms have been conducted to evaluate the proposed algorithms. Furthermore, the best trade-off between the coding efficiency of the proposed error resilience algorithms and error resilience performance has been considered at the design stage. The experimental work evaluation includes both encoding conditions, i.e. error-free and error-prone. The results achieved from the experiments show significant improvements, in (Y-PSNR) results and subjective quality of the decoded bitstream, using the proposed algorithm in error-prone conditions with a variety of packet loss rates. Moreover, experimental work is conducted to test the algorithms complexity in terms of required processing execution time at both encoding and decoding stages. Additionally, the video coding standard performance for both H.264|AVC and H.265|HEVC coding standards are evaluated in error-free and error-prone environments. For ASE algorithm and when compared with improved region of interest (IROI) and region of interest (ROI) algorithms, a significant improvement in visual quality was the most obvious finding from the obtained results with PLRs of 2-18 (%). For ERSEI algorithm and when compared with the default HM16 with pixel copy concealment and motion compensated error concealment (MCEC) techniques, the evaluation results indicate clear visual quality enhancement under different packet loss rates PLRs (1,2 6, 8) %.

Optical fibre has become the dominant theme of transmission in long haul, high data rate communication systems due to its tremendous bandwidth and low loss. Radio over Fibre (RoF) technology facilitates the seamless integration between wireless and optical communication systems and found to be the most promising solution to meet the exponential bandwidth demands expected for the upcoming years. However, the main bit-rate/distance limitation in RoF systems is the chromatic dispersion. In this thesis, the two generations of RoF technologies, namely Analogue RoF (ARoF) and Digital RoF (DRoF) are investigated. The overall aim of this research is to optimise the optical bandwidth utilisation of these two approaches for a typical transmission of the fronthaul link proposed in the next generation Centralised Radio Access Network (C-RAN). Consequently, a number of physical layer design scenarios for the optimised transmission of the Radio Frequency (RF) signals over a Standards Single Mode Fibre (SSMF) are demonstrated. Firstly, for an ARoF transmission, where the analogue RF signals are transported over SSMF using an optical carrier, a bidirectional link transmitting four Downlink/Uplink channels in a chromatic dispersion limited scenario is designed. Simulation results have shown a clear constellation diagram of a 2.5 Gb/s RF signal transmission over 120 km fibre length. Secondly, a DRoF system with reduced optical bandwidth occupancy is proposed. This system employs an optical Duobinary transmission to the digitised RF signal at the transmitter side to reduce its spectrum and to address the chromatic dispersion effect, simultaneously. Simulation results demonstrate the capability of the proposed system to maintain high-quality transmission of the digitised signals over 70 km of fibre distance without dispersion compensation requirements. Finally, an advanced DRoF transmission link based on integrating digital Optical Single Sideband (OSSB) transmission with Duobinary encoding scheme is designed. Simulation results have clearly verified system's robustness against transmission impairments and have better performances in terms of the obtained BER and EVM with respect to the 3GPP standardised values. Moreover, the results show that both transmission distance and power budget are furtherly improved in comparison with two other digital transmission scenarios.

Signal processing is a large subject with applications integral to a number of technological fields such as communication, audio, Voice over IP (VoIP), pattern recognition, sonar, radar, ultrasound and medical imaging. Techniques exist for the analysis, modelling, extraction, recognition and synthesis of signals of interest. The focus of this thesis is signal processing for acoustics (both sonic and ultrasonic). In the applications examined, signals of interest are usually incomplete, distorted and/or noisy. Therefore, reconstructing the signal, noise reduction and removal of any distortion/interference are the main goals of the signal processing techniques presented. The primary aim is to study and develop an advanced time-frequency signal processing technique for acoustic applications to enhance the quality of the signals. In the first part of the thesis, a technique is presented that models and maintains the correlation between temporal and spectral parameters of audio signals. A novel Packet Loss Concealment (PLC) method is developed with applications to VoIP, audio broadcasting, and streaming. The problem of modelling the time-varying frequency spectrum in the context of PLC is addressed, and a novel solution is proposed for tracking and using the temporal motion of spectral flow to reconstruct the signal. The proposed method utilises a Time-Frequency Motion (TFM) matrix representation of the audio signal, where each frequency is tagged with a motion vector estimate that is assessed by cross-correlation of the movement of spectral energy within sub-bands across time frames. The missing packets are estimated using extrapolation or interpolation algorithms using a TFM matrix and then inverse transformed to the time-domain for reconstruction of the signal. The proposed method is compared with conventional approaches using objective Performance Evaluation of Speech Quality (PESQ), and subjective Mean Opinion Scores (MOS) in a range of packet loss from 5% to 20%. The evaluation results demonstrate that the proposed algorithm substantially improves performance by an average of 2.85% and 5.9% in terms of PESQ and MOS respectively. In the second part of the thesis, the proposed method is extended and modified to address challenges of excessive coherent noise arising from ultrasonic signals gathered during Guided Wave Testing (GWT). It is an advanced Non-destructive testing technique which is used over several branches of industry to inspect large structures for defects where the structural integrity is of concern. In such systems, signal interpretation can often be challenging due to the multi-modal and dispersive propagation of Ultrasonic Guided Waves (UGWs). The multi-modal and dispersive nature of the received signals hampers the ability to detect defects in a given structure. The Split-Spectrum Processing (SSP) method with application for such signal has been studied and reviewed quantitatively to measure the enhancement in terms of Signal-to-Noise Ratio (SNR) and spatial resolution. In this thesis, the influence of SSP filter bank parameters on these signals is studied and optimised to improve SNR and spatial resolution considerably. The proposed method is compared analytically and experimentally with conventional approaches. The proposed SSP algorithm substantially improves SNR by an average of 30dB. The conclusions reached in this thesis will contribute to the progression of the GWT technique through considerable improvement in defect detection capability.

Enhancing the separated singing voices from harmonic (pitched) and percussive musical instruments in songs recorded with a single microphone is the scope of this thesis. Separating singing voice has applications in music information retrieval systems. Various methods have been used to separate singing voice from harmonic and percussive instruments. Most of them use two stages of separation, one for separating harmonic instruments, and the other for separating percussive instruments. One of these Algorithms uses non-negative matrix factorization in each stage to separate harmonic and percussive instruments. Traditionally, in each stage, components' bases or gains are clustered based on discontinuity measures. The first contribution of this thesis was the use of local discontinuity of significant parts of these bases and gains, followed by splitting (rather than classifying) each component's basis or gain. This significantly refined the separated voice and music sources. Median filtering has also been used in two stages to separate singing voice. Typically, horizontal and vertical filters are used in each stage. The second contribution of this thesis was to enhance the separation quality using a combination of six additional diagonal median filters to accommodate singing voice frequency modulations. In addition, filters parameters that are suitable for all songs regardless of their sampling frequencies are sought. The third contribution of this research was the novel use of Hough Transform to detect traces of pitched instruments in the magnitude spectrogram of the separated voice. These traces are then removed completely using median filtering after successfully calculating their frequency bands. The new Hough Transform based approach was applied to a number of separation algorithms as a post processing step and it significantly improved the quality of the separated voice and music in all of them.