The Challenge of Vision

Vision systems developed to date do not have the ability of humans in terms of understanding complex indoor or outdoor scenes. Successful machine vision systems need to have a highly constrained environment.

Why is vision hard?

• The world is three-dimensional. Images formed in a camera are two-dimensional projections -- some information has been lost and is difficult to recover.
• The image is dependent on many factors:
  • The nature of the sensing device -- vidicon, CCD, infra-red, greyscale, colour.
  • Properties of the device -- sensitivity, resolution, lenses, stability, focus.
  • The lighting of the scene -- poor lighting gives poor contrast, shadows or excessive reflections may also give problems.
  • The environment -- dust, fog, humidity.
  • The reflective properties of the object surfaces -- texture, colour, specularity.

• A large volume of information has to processed. Whilst this is good in general many vision algorithms require high performance computations.
• Much knowledge about the world - both objects and environment - might need to be represented for vision algorithms.

What sort of operations does a vision system need to perform and why is this relevant to AI?

• Describe images, objects and physical world -- clearly we need (mathematical) models of the image and object and need some knowledge representation scheme as well.
• Image processing -- improve images for human or computer consumption. Highlight/extract relevant features.
• Segmentation -- extract features such as edge, regions, surfaces and other descriptions from an image.
• Pattern recognition -- for single object images, classify the image.
• Measurement analysis -- measure features on the object.
• Image understanding - for multiple object images locate objects in the image, classify them and perhaps build three-dimensional model of the scene.

The Ultimate aim of a vision system is to recognise objects with in an image.

Clearly many AI techniques -- knowledge representation, reasoning, understanding, planning and learning -- are required in the visual process