1 Semantic-Based Workflow Composition for Video Processing in the Grid Gayathri Nadarajan, Yun-Heh Chen-Burger, James Malone Centre for Intelligent Systems and their Applications School of Informatics University of Edinburgh Scotland

2 Introduction - EcoGrid Utilises state-of-the-art Grid technologies to establish an infrastructure for ecological research in several locations in Taiwan. Scientists can conduct data acquisition, data analysis and data sharing on this platform. Divided into four components; network, data streaming, data management and workflow enactment. Collaboration between National Centre for High- Performance Computing (NCHC), Taiwan and Artificial Intelligence Applications Institute (AIAI), University of Edinburgh on workflow enactment for EcoGrid.

3 Sample images from video clips

4 Ecological and Technical Motivations Problem: - Vast amount of valuable and continuous real-time data of varying quality is collected over 3 years, unanalysed. - Manual processing is time-consuming. 1 minute’s clip will take 15 minutes to analyse, thus 1 year’s recording will cost human experts 15 years’ effort x 3 under water cameras = 45 years. Impractical situation and more appropriate automation methods must be deployed. Semi-automated and automated way to speed up this process.

5 Requirements Process Automation: - Process model for video annotation: Performance-Based Selection. Iterative Processing. Adaptive, Flexible and Generic Architecture. Semantic-Based Compatibility. Key frame detection SegmentationRecognitionAnnotation

6 Related Work Existing Grid workflow systems: - Pegasus - Triana - Taverna - Kepler - Askalon - JOpera Limitations of current solutions for flexible workflow composition for complex video processing tasks.

7 Proposed Framework – Hybrid Method

8 Example “Detect at least one tiger fish in high brightness” Goal: Detection Constraints: Occurrence = at_least_one; Brightness = high Process Library Representation: detect_presence(X) :- pre_process(X), feature_extraction(X), segmentation(X), classification(X). pre_process(tiger_fish) :- keyframe_detection, enhancement. Constraint Qualifier in Goal Ontology Illumination_level in Domain Ontology

9 Walkthrough I User (ecologist) interacts with system by providing high level goals and constraints as input. Goal will match that in the Goal ontology. Constraints will match those in the Domain ontology and will determine the video clip to be fetched. The Planner will look for a sequence of processes for execution using task decomposition. Sequence of processes constitute the workflow composition. Instances of processes and their subtasks are contained in the process library.

10 Walkthrough II Each subtask will be matched with a video/image processing operation contained in the Capability ontology. The tool to execute each operation is also contained in the Capability ontology. In cases where more than one tool is available to perform the subtask, the tool with the best performing capability is selected. A set of tools to perform VP task is now obtained. VP tools may now work on the video clip directly.

11 Video Processing Tools/Libraries Open Computer Vision Library (OpenCV) LTI library (C++) HTK (HMM classifier) Lib SVM ITK (medical imaging) STL (C++) Readily available but not Grid-compatible.

12 Discussion Work in progress - Understand video processing tools available. - Build goal, domain and capability ontologies. - Keep terminologies non-technical for ecologists but may have technical implication in underlying mechanism, e.g. “pitch dark” means “brightness” level in the range Build a prototype based on walkthrough

13 Contributions Grid workflow technology with semantic capabilities + AI Planning and CBR. Vision - specialised tools used in workflow could help provide reasonable solutions for automatic video processing, e.g. annotation.

14 End of Slides Thank you!

15 References EcoGrid, National Centre for High Performance, Taiwan, Workflow enactment in the EcoGrid, AIAI, University of Edinburgh, Y-H Chen-Burger, F-P Lin. “A Semantic Based Workflow Choreography for Integrated Sensing and Processing”. In Procs. CNNA G. Nadarajan, Y-H Chen-Burger, J Malone. “Semantic Grid Services for Video Analysis”. In Procs. SerComp06. D. D. O’Roure, N. R. Jennings, N. R. Shaldbolt. “The Semantic Grid: Past, Present and Future”. In Procs. the IEEE Vol. 93, Issue 3, J. Yu, R. Buyya. “A Taxonomy of Workflow Systems for Grid Computing”. Journal of Grid Computing, OpenCV.

16 Proposed Framework – Design Layer Process Manager is responsible for composing sequence of processes to be executed based on available tools. - Planning component - CBR component Ontologies give meaning to process and keep goals separate from capabilities. - Goal - Domain - Tool/Capability

17 Proposed Framework – Workflow Layer Main interface between design and processing layers. Workflow enactor acts as interpreter of events that occur within system.

18 Proposed Framework – Processing Layer Consists of a set of image and video processing tools. Functions of tools represented in capability ontology in Design Layer. Once a workflow has been established, tools may work directly on videos. Depending on quality of video and task at hand, each tool will work on varying level of accuracy. Feedback from domain expert would be useful. Use of ML techniques to assist with performance measure predictions for image and video processing tools.

19 The Semantic Grid The Grid is aimed at enabling resource sharing and coordinated problem-solving between computers and people in a distributed and heterogeneous manner. The Semantic Grid is an extension of current Grid whereby information and services are given well- defined and explicitly represented meaning, i.e. the application of Semantic Web technologies to the Grid. Requires means for facilitating the composition of multiple resources and mechanisms for creating and enacting these in a distributed manner, i.e. composing and executing complex workflows. Improvement of Grid workflow systems with the incorporation of semantic capabilities.