improving object recognition performance through semantic context extraction

Open Access
Smith, Brigid Louise
Area of Honors:
Computer Engineering
Bachelor of Science
Document Type:
Thesis Supervisors:
  • Vijaykrishnan Narayanan, Thesis Supervisor
  • Lee David Coraor, Honors Advisor
  • computer vision
  • machine vision
  • vision
  • context extraction
  • image processing
Machine vision is a computationally expensive problem with an exceptionally large number of real-world applications. With the rise of the Internet of Things and the presence of wearables in day to day settings, there is an additional focus on power constraints and the limitations of fixed hardware. In a vision pipeline, the accuracy of the object classification stage will likely affect the usefulness of the pipeline as a whole. However, we find that it is difficult to create a system with the ability to recognize a large number of objects both quickly and accurately because the number of classifiers needed grows with the number of objects. We observe that real world images and the objects in them tend to be sensible and expose relationships between objects and scenes that are used by humans intuitively. This high-level context could potentially be used to inform and improve object classification by allowing us to make reasonable, probabilistic guesses about objects that might occur based on other information that we have about the image. This guesswork will lower the number of classifiers that need to be run, which will also address power and timing concerns. In this paper, we explore the meaning of context, design a framework to store it in a way accessible to a computer, and then evaluate the efficacy of context-based filtering.