Sue Inn Ch'ng

A new multi-purpose audio-visual UNMC-VIER database with multiple variabilities

Audio-visual recognition system is becoming popular because it overcomes certain problems of traditional audio-only recognition system. However, difficulties due to visual variations in video sequence can significantly degrade the recognition performance of the system. This problem can be further complicated when more than one visual variation happen at the same time. Although several databases have been created in this area, none of them includes realistic visual variations in video sequence. With the aim to facilitate the development of robust audio-visual recognition systems, the new audio-visual UNMC-VIER database is created. This database contains various visual variations including illumination, facial expression, head pose, and image resolution variations. The most unique aspect of this database is that it includes more than one visual variation in the same video recording. For the audio part, the utterances are spoken in slow and normal speech pace to improve the learning process of audio-visual speech recognition system. Hence, this database is useful for the development of robust audio-visual person, speech recognition and face recognition systems.

Uninformed pathfinding

Proposal of the boundary iterative-deepening depth-first search (BIDDFS) algorithm.The BIDDFS is extended for bidirectional searching - the bidirectional BIDDFS.A parallel approach is applied to the bidirectional BIDDFS.The BIDDFS is enhanced to search for multiple goals - the multi-goal BIDDFS.Simulations showed time improvements for the proposed uninformed algorithms. This paper presents a new pathfinding algorithm called the boundary iterative-deepening depth-first search (BIDDFS) algorithm. The BIDDFS compromises the increasing memory usage of the Dijkstras algorithm, where the memory clears enables the BIDDFS to consume less memory than the Dijkstras algorithm. The expansion redundancy of the iterative-deepening depth-first search (IDDFS) is also compensated; it is faster than the IDDFS in all of the testing instances conducted. The BIDDFS is further enhanced for bidirectional searching to allow expanding to fewer nodes and reducing pathfinding time. The bidirectional BIDDFS and the parallel bidirectional BIDDFS are also proposed. The proposed BIDDFS is further extended to the multi-goal BIDDFS, which is able to search for multiple goals present on the map in a single search. Simulation examples and comparisons have revealed the good performance of the proposed algorithms.

Modular dynamic RBF neural network for face recognition

Over the years, we have seen an increase in the use of RBF neural networks for the task of face recognition. However, the use of second order algorithms as the learning algorithm for all the adjustable parameters in such networks are rare due to the high computational complexity of the calculation of the Jacobian and Hessian matrix. Hence, in this paper, we propose a modular structural training architecture to adapt the Levenberg-Marquardt based RBF neural network for the application of face recognition. In addition to the proposal of the modular structural training architecture, we have also investigated the use of different front-end processors to reduce the dimension size of the feature vectors prior to its application to the LM-based RBF neural network. The investigative study was done on three standard face databases; ORL, Yale and AR databases.

Multiview-multiband face recognition system to solve illumination and pose variation

Identifying faces under the influence of illumination and pose can be challenging as the presence of two variations on the same image can greatly change the appearance of a person. Thus, in this paper, we propose a multiview face recognition system that is able to solve illumination and pose face recognition problems. The proposed system uses multiband feature technique to extract features that are invariant to illumination variation and parallel radial basis function neural networks to train different poses. The recognition performance of the proposed system is validated against the Yale B database and compared to other systems implemented on the same database.

Enhanced multiband feature technique for face recognition under varying illumination

This paper presents an enhanced multiband feature technique to improve the performance of face recognition under varying illumination. First, the illumination invariant subbands are extracted using wavelet packet transform and multiband feature selector. Then, histogram equalization is applied to the selected subbands to enhance the contrast of the subband (global). To reduce the noise and enhance the fine details of the facial features (local), an unsharp filter is subsequently applied to the histogram equalized subband. The unsharp filter is created by combining a Gaussian low pass filter and a negative Laplacian operator. The recognition performance of the proposed enhancement scheme is validated against the Yale B database. An improvement in recognition rate has been observed when the enhancement scheme is compared to the original unenhanced subband.

Pathfinding for the navigation of visually impaired people

A navigation system using an Android mobile device for the visually impaired is explored in this paper. This paper focuses on pathfinding algorithms and their implementations on Java platform. The boundary iterative-deepening depth-first search (BIDDFS) pathfinding algorithm is extended for bidirectional searching. Fast pathfinding is applied for the BIDDFS by reducing memory read and writes cycles, proposing the optimised BIDDFS. Fast pathfinding is also extended for the bidirectional BIDDFS, proposing the fast bidirectional BIDDFS. The fast bidirectional BIDDFS uses Javas thread feature to implement a parallel structure. The optimised BIDDFS was able to record drastic improvements in pathfinding speeds compared to the standard BIDDFS. Likewise, the fast bidirectional BIDDFS recorded significant speed improvements over the parallel bidirectional BIDDFS.

Curvelet-based illumination invariant feature extraction for face recognition

This paper presents a curvelet-based illumination invariant feature extraction technique to solve the problem of varying illumination in face recognition. Multiband feature technique is employed to search the decomposed curvelet subbands for subbands which are insensitive to illumination variation. The two best performing subbands are then concatenated to form the Optimal Curvelet Subbands (OCS). To further improve the performance of OSC, histogram equalization is applied to enhance the contrast of the details. The proposed feature extraction method was evaluated on YaleB, EYaleB and AR database. The simulation results obtained shows that the proposed method outperforms its wavelet counterpart and that the extracted subbands are also applicable for other databases.

Trip planning route optimization with operating hour and duration of stay constraints

This paper proposes an offline approach to generating an optimal vacation routing plan to maximize the number of places to visit during the vacation duration. A traveling salesman problem algorithm results in a plan with the shortest path of travel given a number of stops. This solution is not suitable as a travel planner needs to take into account the operating hours and duration of stay at each stop to create a daily plan. The proposed approach will calculate the time of arrival and minimum time of stay to determine if exceeds the operating hours of the stop. The most important criteria of this system is the need to calculate the time of travel from each location. This is done using two methods, 1) using the latitudinal and longitudinal coordinates of places to calculate the straight-line distance and subsequently time between them, and 2) by constructing a look-up table consisting of the travel distance and travel time between places using the Google Maps Directions API. The simulation results show that the total traveling time of the proposed system is within 15 minutes to online results which is optimized using data from the Google Maps Directions API while being able to satisfy both operating hours and duration of stay constraints.

Online and offline electronic question and answer system: Quick question

Asian students with conservative behaviors tend to feel embarrassed and reluctant to ask questions in large classrooms. This paper proposed a mobile-compatible web-based question and answer application to address the aforementioned problem. This proposed web application is known as Quick Question (QQ). QQ is designed to allow students to post questions anonymously and rapidly during class hours. QQ can be accessed online from a web hosting server, or offline from a Raspberry Pi without Internet access. QQ can run on multiple platforms with different screen sizes and capable of serving 250 concurrent users while using the Raspberry Pi. QQ also includes a profanity filter that censor bad words with an accuracy of 86.11%. Furthermore, a term extraction and learning mechanism is implemented to help in grouping and sorting questions posted by students in a class.

The effect of using super-resolution to improve feature extraction and registration of low resolution images in sensor networks

In this paper, the effect of using multi-image and single-image super-resolution to reduce registration errors of low resolution images is evaluated. Two sets of low resolution images were captured using CMUCam4 to perform the evaluation. Moreover, a simplified method that make use of feature points extracted from resolution enhanced / upscaled images to improve the registration of low resolution images is also presented. The simulation results show that enhancing / upscaling the images in prior to registration does help to reduce the registration errors.