|
Digital Processing of Quantum-Limited Images Conjecture on the Relationship between Spatial and Temporal Visual Processes Why do Stabilized Images Disappear? A Simple Model for Filling-In, Contrast, Contrast Constancy and Assimilation
|
Digital Processing of Quantum-Limited Images The physical nature of light and its interaction with matter is such that the detection of light occurs as discrete events (the sensing of individual quanta) whose occurances are randomly distributed over time and space. As a result, when a scene reflects or generates very low light levels and the scene is sensed by a device of sensitivity sufficient to detect individual quanta, the image will appear noisy or speckled, each point in the displayed image randomly fluctuating in brightness from instant to instant. This noise is called photon noise, or quantum noise. The most common example of such imagery can be seen in the displays of night vision devices. All sensing devices generate some noise within their structure, over and above the quantum noise in the image. If the internal noise is small compared with the quantum noise in an image, then the image is said to be quantum-limited.
When the light level is extremely low, quantum noise seriously obscures the content of an image. For example, Figure 1b is a (simulated) quantum-noisy image of the object (a runway) in Figure 1a. A pilot seeing the runway at night through night vision goggles would see an image that looks like that in 1b, except that the speckle pattern would be constantly changing. (This particular figure simulates an image in which an average of 14 photons were sensed per pixel, the image being 100x100 pixels in size.) Figure 1c shows the result of the application of an image-processing algorithm to this image, an algorithm that uses knowledge about the statistics of light to improve the image. Call this "algorithm T". Figure 1d shows the application of a different algorithm that also makes use of knowledge of the physics of light. Call this "algorithm S". Figure 1e shows the results of the application of both T and S algorithms. Figure 1f shows the results of the application of different algorithm, call it "algorithm V", that is based upon a model of the human visual system. Finally, Figure 1g is the result of applying all three algorithms to the image. Note that, while the image in 1g looks fuzzy, the object (runway) is very much more detectable, and in fact, because of the speckle, the edges of the runway in the unprocessed image, Figure 1b, are really no more sharply defined that those in Figure 1g. That is, the application of these algorithms vastly improves the detectability of the object without any penalty (except the requirement that computations must be made). |
