Images

The issues of uncertainty about images and ambiguity about the nature or interpretation of the scenes depicted in them are enormously problematic in image processing which is a complex discipline involving image analysis, shape analysis, texture analysis, feature extraction, digital image correlation, image segmentation and recognition, and morphometrics, including both landmark-based and outline-based methods.

Quantitative image analysis must cope with uncertainty in images arising from

• • Poor pixel or colour resolution,

  • Registration errors, missing calibration or scaling references,

  • Blurring, low bandwidth, de-noising strategies,

  • Missing imagery from data loss or corruption or obscuration by clouds,

  • Ambiguity about landmarks and what appears in the image,

  • Deformation of non-rigid objects during image capture,

  • Missing or errant sensors in image capture,

• Possible errors or overzealousness in image enhancement (“cleaning”),

  • Mistakes in digital image correlation (DIC) or image correction using stereoscopic, tomographic, geometric or similar methods.

The image below is a distorted detail from The Unicorn Is Found, one of the famous medieval Unicorn Tapestries, indicative of the result of compositing sequential images of the tapestry whose threads were relaxing and deforming during the raster-like image capture process. Where is the third man's left eye?

Beyond the problems that can be characterised quantitatively, are the even more subtle issues that arise when imagery is used heuristically, qualitatively or semi-qualitatively in medical diagnosis, forensic argumentation, facial recognition and other applications.

Prof Kenneth Shroyer, chair of pathology at Stony Brook University, explains that distinguishing cancerous cells in a histopathology image is like recognizing your Aunt Millie. It is hard to explain how you know it is her, but you can usually pick her out of a crowd. At the same time, sometimes there is uncertainty about whether it actually Aunt Millie, or whether the cells are cancerous, bwcause of the lighting or the angle or things in the way or just because Millie and cancerous cells sometimes look different from day to day. Pathologists have no standardized schemes for quantifying, reporting or even acknowledging these kinds of uncertainties, except to request another image be taken. It is generally unclear whether or how much such uncertainties are relevant in medical diagnosis.