My research focuses on advanced image segmentation algorithms, including parameter-free hierarchical methods and LiDAR data processing for building extraction. I emphasise non-iterative, flexible approaches that improve image segmentation quality and efficiency. Key contributions include developing hierarchical mutual nearest neighbour and tree-based segmentation techniques, and innovative methods for automatic building roof extraction from LiDAR data. Additionally, my work extends to analysing human interactions in meetings using directed acyclic graphs and developing new measures for correlation mining in graph databases.
SM Abdullah, P Tischer, S Wijewickrema, A Paplinski
2017 IEEE Visual Communications and Image Processing (VCIP), 2017
This paper presents a novel, parameter-free hierarchical image segmentation algorithm that uses minimum spanning trees. The method converts an image into a tree structure and generates multiple levels of segmentation in a single pass, without the need for iterative processing. It considers both segment levels and distances when merging, resulting in more homogeneous segments. The algorithm is versatile, as it can work with any distance function and can be adapted for clustering. A key feature is its segment visualizer, which allows users to select the most appropriate level of segmentation for their needs. When tested on popular segmentation datasets, the method showed competitive performance compared to existing algorithms, while offering the advantages of being parameter-free and non-iterative. This approach provides a quick and flexible way to explore the natural connectivity of image pixels at different levels of detail.
SM Abdullah, P Tischer, S Wijewickrema, A Paplinski
2016 International Conference on Digital Image Computing: Techniques and Applications (DICTA), 2016
This paper presents a novel approach to image segmentation called the Hierarchical Mutual Nearest Neighbour (HMNN) method. Unlike many existing algorithms that rely on thresholds or parameter settings, this method is parameter-free and generates a hierarchy of segmentations at different levels of detail. It works by first creating natural segments based on the principle of mutual nearest neighbors, then merging these segments hierarchically. The algorithm uses only one meta-parameter (k) to control the level of inclusion. The method was evaluated using the Berkeley BSD500 dataset and compared favorably with existing hierarchical segmentation algorithms, outperforming them in several metrics. The HMNN method is versatile, easily applicable to different types of images, and can be extended to handle higher-dimensional images. Its simplicity and effectiveness make it a promising approach for various computer vision applications.
SM Abdullah, M Awrangjeb, G Lu
2014 IEEE international conference on multimedia and expo workshops (ICMEW), 1-6, 2014
This paper presents a new method for automatically extracting building roofs from LiDAR (Light Detection and Ranging) data. The technique first separates ground and non-ground points, then analyzes the non-ground points at different height levels to identify coplanar points that could form roof segments. It uses a region-growing approach to expand these segments into full roof planes. The method then applies rules to remove false positives, such as planes detected on trees or other non-building structures. Finally, it combines the remaining roof planes to determine building outlines. The approach was tested on benchmark datasets from Vaihingen, Germany, and showed high accuracy in both building detection and roof plane extraction, outperforming a previous method in several metrics. This technique is significant because it uses only LiDAR data, avoiding the complexities of integrating image data, while still achieving robust results in identifying buildings and their roof structures.
SM Abdullah, M Awrangjeb, G Lu
ISPRS Technical Commission III Symposium, 2014
This paper presents a new method for automatically detecting buildings and extracting roof planes from LiDAR (Light Detection and Ranging) point cloud data. The method first separates ground and non-ground points, then analyzes the coplanarity of non-ground points to identify potential roof surfaces. It uses an innovative approach to select seed points for growing planar segments, starting from the highest points and working downwards. A rule-based system is then applied to remove segments likely to be on trees rather than buildings. The method was tested on six different datasets with varying terrain and vegetation density. Results show high accuracy in both building detection and roof plane extraction, even in areas with dense vegetation. The approach outperforms a previous method in most metrics, particularly in the accuracy of building boundaries. This technique could be valuable for applications like urban planning, 3D city modeling, and disaster management.
A Fariha, CF Ahmed, CKS Leung, SM Abdullah, L Cao
Pacific-Asia Conference on Knowledge Discovery and Data Mining, 38-49, 2013
This research paper introduces a new method for analyzing and understanding human interactions in meetings using directed acyclic graphs (DAGs). Unlike previous tree-based approaches, the DAG-based method captures both temporal and triggering relationships between interactions, providing a more comprehensive representation of meeting dynamics. The authors propose a weighted DAG model where each interaction is assigned a weight based on the rank or importance of the person initiating it. They also develop an algorithm called WDAGmeet to mine frequent interaction patterns from these weighted DAGs. The method distinguishes between interactions initiated by people of different ranks and captures patterns that might be missed by tree-based approaches. Experimental results show that the WDAGmeet algorithm effectively discovers meaningful interaction patterns from meeting data, potentially offering insights into meeting effectiveness and decision-making processes.
M Samiullah, CF Ahmed, MA Nishi, A Fariha, SM Abdullah, MR Islam
Web Technologies and Applications: 15th Asia-Pacific Web Conference, 2013
This paper introduces a new approach to correlation mining in graph databases, addressing the challenge of identifying meaningful relationships between graph structures. The authors propose a novel measure called gConfidence and an algorithm named CGM (Correlated Graph Mining) to efficiently extract correlated graph patterns. Unlike existing methods that focus on structural similarity, this approach can identify correlations between graphs that appear together frequently, even if they are structurally dissimilar. The gConfidence measure has useful properties such as downward closure, which allows for efficient pruning of the search space. The CGM algorithm constructs a hierarchical search space called a gConfidence tree to mine correlations effectively. Experimental results on both synthetic and real-world datasets demonstrate that CGM outperforms existing algorithms in terms of processing speed and ability to filter out less significant graphs. This method has potential applications in various domains, including bioinformatics, computer vision, and chemical analysis.
M Samiullah, SM Abdullah, AFMIH Bappi, S Anwar
2012 International Conference on Informatics, Electronics & Vision (ICIEV), 2012
This paper proposes a new congestion control protocol for Wireless Body Sensor Networks (WBSNs) used in healthcare monitoring. The protocol aims to improve reliability and energy efficiency by managing queue occupancy in sensor nodes. It introduces a Queue Status (QS) bit to indicate congestion levels and uses a backpressure mechanism to adjust packet sending rates. The protocol detects congestion by monitoring queue lengths at nodes and their successors. When congestion is detected, it sends backpressure messages to slow down upstream nodes. Simulation results show the protocol can reduce packet drop rates, improve delivery ratios, and decrease energy consumption as buffer sizes increase. The authors claim their approach provides an effective way to control congestion in WBSNs while considering the limited computational and power resources of sensor nodes.