Information-Conserving Object Recognition,
Abstract
Charge-coupled device (CCD) cameras typically produce scene images with extremely low but nonzero noise variance. In fact, for object recognition purposes in computer vision, an initial assumption often is that the noise can be neglected so that the data at each pixel can be regarded as deterministic. In the present investigation, however, we take an alternative approach that follows a strictly physical interpretation of classical estimation theory. First, we use experimental data to determine the joint probability distribution of the pixel brightness measurements in our CCD images. We use this to construct the likelihood function for any parameter set that is to be estimated given our image data. It is significant that the form of the likelihood function in this physical approach is not arbitrary, but depends upon the probability distribution of the brightness measurements no matter how low the corresponding noise variance is at each pixel, as long as it is nonzero. Moreover, it is the form of this likelihood function, not the level of the noise, that determines the optimal method of recognizing an imaged object.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 1997
- Accession Number
- ADA353673
Entities
People
- Margrit Betke
- Nicholas C. Makris
Organizations
- University of Maryland