Identification of an individual using physical attributes such as speech for example is of increa... more Identification of an individual using physical attributes such as speech for example is of increasing commercial importance in the field of security. Speech recognition is one method of authenticating an individual based on extracting features from spoken words and then classifying them as belonging to a particular person. This technique has good results when used in acoustically noise free environments but has limited success in busy environments such as offices, airports, train stations, factory floors etc or in the presence of multiple talkers. Recent research has focussed on the use of visual information to supplement acoustic speech which results in a more robust, noise-tolerant system[4]. The identification and visual tracking of lips is commonly employed. However, little work has investigated the extraction of lip features in natural surroundings, and the best way of synchronising and integrating lip movement with speech recognition.
TENCON 2021 - 2021 IEEE Region 10 Conference (TENCON), Dec 7, 2021
The design and optimization of piezoelectric transducer (PT) are one of the most challenging and ... more The design and optimization of piezoelectric transducer (PT) are one of the most challenging and critical stages in acoustic power transfer (APT). In this paper, the equivalent circuit of the PT is obtained by electromechanical analogy, and it is found that the system efficiency has nothing to do with the shape of the PT. However, the volume parameters (sectional area and thickness) of the PT greatly influence the efficiency of the system. We found that the efficiency of the system increases as the cross-sectional area of the PT increases. The thickness is just the opposite, and the system efficiency decreases as the thickness increases. In order to verify this law, the paper also established a system model in COMSOL and found that the results are highly consistent with the theoretical model. Optimizing the parameters of the PT can reduce the weight of the PT applied in life and broaden the application range for the use of the PT array.
Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, Dec 7, 2020
This paper presents an integrated architecture of the class embodiment autoencoder (CEAE) and var... more This paper presents an integrated architecture of the class embodiment autoencoder (CEAE) and variational autoencoder. The aim is to improve the generalisation of the algorithm and accordingly increase the classification accuracy of unseen samples. The proposed variational CEAE is trained by using hyperspectral images of Manuka honey dataset, then evaluated for generalisation performance on unseen brands of honey. We applied well-known generalisation techniques to this structure, and evaluated the effect of these on our dataset. Our experiment results show that the average validation set performance of the new autoencoder technique on unseen brands is 55.4%, while the average benchmark technique is 48.1% for the same unseen brands. The autoencoder structures are performing feature reduction on our data, which has shown to improve the classification accuracy and generalisation performance. We tested the feature reduction techniques in combination with K-nearest-neighbour classifier, linear support vector machine (SVM), and radial basis function SVM. This work develops an important step toward the automatic classification of Manuka honey quality using hyperspectral imaging and machine learning. This is the first work to evaluate generalisation performance in honey classification, which is crucial for a viable real-world solution.
Information Security Journal: A Global Perspective, Sep 14, 2016
ABSTRACT The explosive growth in fingerprint technologies within the past decade has seen the eme... more ABSTRACT The explosive growth in fingerprint technologies within the past decade has seen the emergence of a dedicated field of research into securing fingerprint templates during storage in a database. While new fingerprint template protection techniques are often broadly classified as belonging to the well-known salting, noninvertible transforms, key binding, or key generation categories, methods within each category are currently lacking a sense of organization. This article aims to fill this gap by proposing a categorization of noninvertible fingerprint transforms based on their design mechanisms. Our survey of the current literature in this field reveals two prominent types of approaches, so we classify existing noninvertible fingerprint transforms into two main categories: perturbation-based and histogram-based. We also discuss the evaluation techniques used to assess the robustness of noninvertible fingerprint transforms in the literature. These contributions will serve to help researchers find their bearing in the growing fingerprint template protection field, thereby encouraging a deeper understanding of the field and faster progress in the development of more effective fingerprint template protection schemes.
2019 International Conference on Electronics, Information, and Communication (ICEIC), 2019
Common issues associated with wireless sensor networks for analysing environmental noise include ... more Common issues associated with wireless sensor networks for analysing environmental noise include high power usage, high cost, and limited scalability. In this paper, we present a novel approach to develop an autonomous system for collecting environmental noise information. We designed a system to be highly scalable, easy to use, low-cost, and low-powered to encourage its widespread adoption. Power usage is kept low by investigating the use of Bluetooth Low Energy communication protocol for streaming environmental noise. Beyond merely detecting noise levels, the proposed system can classify the source of the measured noise. Mel-frequency cepstrum coefficients are used to produce features of the noise and model Gaussian mixture models to classify the source of the noise. Information from the system is sent to a cloud storage service in real-time which enables visualization of noise levels, noise sources and battery level of the sensors in our web application. The system was able to reliably collect information over a 4-week field test. It has an average classification accuracy of 72% when subjected to four common environmental noise sources.
In this paper, we propose a new time-frequency mask method for computational auditory scene analy... more In this paper, we propose a new time-frequency mask method for computational auditory scene analysis (CASA) based on convex optimization of the binary mask. In the proposed method, the pitch estimation and segment segregation in conventional CASA are completely replaced by the convex optimization of speech power. Considering the cross-correlation between the power spectra of noisy speech and noise in each of a Gammatone filterbank channel, the objective function of speech power used for convex optimization is built. The speech power is estimated by gradient descent method. Thus, the time-frequency units dominated by speech and noise are labeled by comparing the powers of noisy and estimated speech, and noise. The erroneous local masks are also removed by using the Teager energy of the estimated speech and time-frequency unit smoothing. The results from the average segmental signal-to-noise ratio improvement, HIT-False Alarm rate and subjective test show that the performance of the proposed method outperforms the reference methods.
Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, Dec 7, 2020
Biometric recognition methods using human traits like fingerprint, face, voice, palm-print, and p... more Biometric recognition methods using human traits like fingerprint, face, voice, palm-print, and palm vein have developed significantly in recent years. Palm vein recognition has gained attention because of its unique characteristics and high recognition accuracy. Many palm vein recognition methods proposed recently suffer from the issue of having low-quality images right at the acquisition stage, resulting in degradation of recognition accuracy. This paper proposes the use of a Convolutional Neural Network (CNN); U-Net, to effectively segment the vein networks from the background of near-infrared palm vein images. The experiments were conducted on the HK PolyU Multispectral Palmprint and Palmvein database. The original images taken from the database were reduced to region of interests. Morphological operations were applied to obtain ground truth mask images. The mask images were then used to train a modified U-Net in which Gabor filter was applied in the first block of the U-Net architecture. The accuracy of the segmented vein images was obtained by determining the overlap between the segmented images obtained from the network and the corresponding ground truth images from the morphological operations. The overlap is evaluated using the Jaccard Index and Dice Coefficient Metrics. For both of these similarity metrics, the value 0” indicates no overlap and 1” indicates a complete congruence between the subject images. The best Dice Coefficient obtained in this experiment is 0.69 and the Jaccard Index is 0.71, which makes this technique promising for automatic vein segmentation and can be adopted in palm vein recognition systems.
Autoencoders have shown to be very useful when applied to preprocessing, and pretraining for neur... more Autoencoders have shown to be very useful when applied to preprocessing, and pretraining for neural networks. They have also been applied to feature extraction, and data compression although they are not widely used in these applications. For feature extraction, a common problem with autoencoders is that they learn features that can reconstruct the data. However, these features are not necessarily sufficient when it comes to classifying the data. Recently Autoencoders have been adapted for use in semi supervised tasks by introducing a new layer as a classification output during the training process. This structure aims to train features that are good at both data classification and reconstruction and are called supervised autoencoders. In this paper, we apply a new variant of this technique called the Class Embodiment Autoencoder to hyperspectral imaging data of honey samples where we aim to classify the botanical origins of New Zealand honey.
Identification of an individual using physical attributes such as speech for example is of increa... more Identification of an individual using physical attributes such as speech for example is of increasing commercial importance in the field of security. Speech recognition is one method of authenticating an individual based on extracting features from spoken words and then classifying them as belonging to a particular person. This technique has good results when used in acoustically noise free environments but has limited success in busy environments such as offices, airports, train stations, factory floors etc or in the presence of multiple talkers. Recent research has focussed on the use of visual information to supplement acoustic speech which results in a more robust, noise-tolerant system[4]. The identification and visual tracking of lips is commonly employed. However, little work has investigated the extraction of lip features in natural surroundings, and the best way of synchronising and integrating lip movement with speech recognition.
TENCON 2021 - 2021 IEEE Region 10 Conference (TENCON), Dec 7, 2021
The design and optimization of piezoelectric transducer (PT) are one of the most challenging and ... more The design and optimization of piezoelectric transducer (PT) are one of the most challenging and critical stages in acoustic power transfer (APT). In this paper, the equivalent circuit of the PT is obtained by electromechanical analogy, and it is found that the system efficiency has nothing to do with the shape of the PT. However, the volume parameters (sectional area and thickness) of the PT greatly influence the efficiency of the system. We found that the efficiency of the system increases as the cross-sectional area of the PT increases. The thickness is just the opposite, and the system efficiency decreases as the thickness increases. In order to verify this law, the paper also established a system model in COMSOL and found that the results are highly consistent with the theoretical model. Optimizing the parameters of the PT can reduce the weight of the PT applied in life and broaden the application range for the use of the PT array.
Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, Dec 7, 2020
This paper presents an integrated architecture of the class embodiment autoencoder (CEAE) and var... more This paper presents an integrated architecture of the class embodiment autoencoder (CEAE) and variational autoencoder. The aim is to improve the generalisation of the algorithm and accordingly increase the classification accuracy of unseen samples. The proposed variational CEAE is trained by using hyperspectral images of Manuka honey dataset, then evaluated for generalisation performance on unseen brands of honey. We applied well-known generalisation techniques to this structure, and evaluated the effect of these on our dataset. Our experiment results show that the average validation set performance of the new autoencoder technique on unseen brands is 55.4%, while the average benchmark technique is 48.1% for the same unseen brands. The autoencoder structures are performing feature reduction on our data, which has shown to improve the classification accuracy and generalisation performance. We tested the feature reduction techniques in combination with K-nearest-neighbour classifier, linear support vector machine (SVM), and radial basis function SVM. This work develops an important step toward the automatic classification of Manuka honey quality using hyperspectral imaging and machine learning. This is the first work to evaluate generalisation performance in honey classification, which is crucial for a viable real-world solution.
Information Security Journal: A Global Perspective, Sep 14, 2016
ABSTRACT The explosive growth in fingerprint technologies within the past decade has seen the eme... more ABSTRACT The explosive growth in fingerprint technologies within the past decade has seen the emergence of a dedicated field of research into securing fingerprint templates during storage in a database. While new fingerprint template protection techniques are often broadly classified as belonging to the well-known salting, noninvertible transforms, key binding, or key generation categories, methods within each category are currently lacking a sense of organization. This article aims to fill this gap by proposing a categorization of noninvertible fingerprint transforms based on their design mechanisms. Our survey of the current literature in this field reveals two prominent types of approaches, so we classify existing noninvertible fingerprint transforms into two main categories: perturbation-based and histogram-based. We also discuss the evaluation techniques used to assess the robustness of noninvertible fingerprint transforms in the literature. These contributions will serve to help researchers find their bearing in the growing fingerprint template protection field, thereby encouraging a deeper understanding of the field and faster progress in the development of more effective fingerprint template protection schemes.
2019 International Conference on Electronics, Information, and Communication (ICEIC), 2019
Common issues associated with wireless sensor networks for analysing environmental noise include ... more Common issues associated with wireless sensor networks for analysing environmental noise include high power usage, high cost, and limited scalability. In this paper, we present a novel approach to develop an autonomous system for collecting environmental noise information. We designed a system to be highly scalable, easy to use, low-cost, and low-powered to encourage its widespread adoption. Power usage is kept low by investigating the use of Bluetooth Low Energy communication protocol for streaming environmental noise. Beyond merely detecting noise levels, the proposed system can classify the source of the measured noise. Mel-frequency cepstrum coefficients are used to produce features of the noise and model Gaussian mixture models to classify the source of the noise. Information from the system is sent to a cloud storage service in real-time which enables visualization of noise levels, noise sources and battery level of the sensors in our web application. The system was able to reliably collect information over a 4-week field test. It has an average classification accuracy of 72% when subjected to four common environmental noise sources.
In this paper, we propose a new time-frequency mask method for computational auditory scene analy... more In this paper, we propose a new time-frequency mask method for computational auditory scene analysis (CASA) based on convex optimization of the binary mask. In the proposed method, the pitch estimation and segment segregation in conventional CASA are completely replaced by the convex optimization of speech power. Considering the cross-correlation between the power spectra of noisy speech and noise in each of a Gammatone filterbank channel, the objective function of speech power used for convex optimization is built. The speech power is estimated by gradient descent method. Thus, the time-frequency units dominated by speech and noise are labeled by comparing the powers of noisy and estimated speech, and noise. The erroneous local masks are also removed by using the Teager energy of the estimated speech and time-frequency unit smoothing. The results from the average segmental signal-to-noise ratio improvement, HIT-False Alarm rate and subjective test show that the performance of the proposed method outperforms the reference methods.
Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, Dec 7, 2020
Biometric recognition methods using human traits like fingerprint, face, voice, palm-print, and p... more Biometric recognition methods using human traits like fingerprint, face, voice, palm-print, and palm vein have developed significantly in recent years. Palm vein recognition has gained attention because of its unique characteristics and high recognition accuracy. Many palm vein recognition methods proposed recently suffer from the issue of having low-quality images right at the acquisition stage, resulting in degradation of recognition accuracy. This paper proposes the use of a Convolutional Neural Network (CNN); U-Net, to effectively segment the vein networks from the background of near-infrared palm vein images. The experiments were conducted on the HK PolyU Multispectral Palmprint and Palmvein database. The original images taken from the database were reduced to region of interests. Morphological operations were applied to obtain ground truth mask images. The mask images were then used to train a modified U-Net in which Gabor filter was applied in the first block of the U-Net architecture. The accuracy of the segmented vein images was obtained by determining the overlap between the segmented images obtained from the network and the corresponding ground truth images from the morphological operations. The overlap is evaluated using the Jaccard Index and Dice Coefficient Metrics. For both of these similarity metrics, the value 0” indicates no overlap and 1” indicates a complete congruence between the subject images. The best Dice Coefficient obtained in this experiment is 0.69 and the Jaccard Index is 0.71, which makes this technique promising for automatic vein segmentation and can be adopted in palm vein recognition systems.
Autoencoders have shown to be very useful when applied to preprocessing, and pretraining for neur... more Autoencoders have shown to be very useful when applied to preprocessing, and pretraining for neural networks. They have also been applied to feature extraction, and data compression although they are not widely used in these applications. For feature extraction, a common problem with autoencoders is that they learn features that can reconstruct the data. However, these features are not necessarily sufficient when it comes to classifying the data. Recently Autoencoders have been adapted for use in semi supervised tasks by introducing a new layer as a classification output during the training process. This structure aims to train features that are good at both data classification and reconstruction and are called supervised autoencoders. In this paper, we apply a new variant of this technique called the Class Embodiment Autoencoder to hyperspectral imaging data of honey samples where we aim to classify the botanical origins of New Zealand honey.
Uploads
Papers by Waleed Abdulla