Thesis
Scalable Continuous Learning for Visual Object Detection
Southern Cross University
Masters by Thesis, Southern Cross University
2020
DOI:
https://doi.org/10.25918/thesis.299
Metrics
6 File views/ downloads
37 Record Views
Abstract
Object detection is a subset of artificial intelligence that enables machines to automatically segment and classify objects in an image or video based on visual appearance. The deep learning approach uses a large quantity of data to autonomously capture the scope of object classes and samples per class. The use of big data raises a scalability issue in deep learning, which is the need for long computation time, a large amount of storage to store the data, and a large random-access memory (RAM) to process the image data in batches. Another scalability issue is the need to support continuous learning for expanding the scope of classifiable object classes and adding more samples for the existing classes to improve the robustness and accuracy.
This thesis investigates how continuous learning can be accomplished by a knowledge transfer method to fine-tune the existing deep learning models using only new training data. The experimental results demonstrate that the knowledge transfer method can support a more scalable deep learning that allows for better retention of learned knowledge without the need to reprocess past data, which consequently reduces the requirement for storage and computing memory while increasing the processing time. Thus, this thesis provides both theoretical and practical contributions to promote the future use of knowledge transfer for scalable, deep learning, real-world object detection applications that naturally need to continuously expand the scope of classifiable objects and increase the accuracy by increasing the samples per class. Further, knowledge transfer will also support faster fine-tuning of generic object detection models for domain-specific deployments.
Details
- Title
- Scalable Continuous Learning for Visual Object Detection
- Creators
- Elizabeth Irenne Yuwono
- Contributors
- Golam Sorwar (Supervisor) - Southern Cross UniversityDian Tjondronegoro (Supervisor) - Griffith UniversityAlireza Alaei (Supervisor) - Southern Cross University
- Awarding Institution
- Southern Cross University; Masters by Thesis
- Theses
- Masters by Thesis, Southern Cross University
- Publisher
- Southern Cross University
- Number of pages
- xii, 134
- Identifiers
- 991013139112402368
- Copyright
- © EI Yuwono 2020
- Academic Unit
- Faculty of Business, Law and Arts; School of Business and Tourism; Information Technology
- Resource Type
- Thesis