On Monday — just before staffing The MathWorks’ booth — I attended Maria Petrou’s plenary session: “Reverse-Engineering the Human Visual System.” I found it rather interesting, though it tried to cover too much ground for one hour.* Nevertheless, it was quite interesting to ponder how we can use our understanding of the human visual system to perform better digital image processing.
Among the more interesting ideas:
- There’s a difference between vision and perception. One concerns stimuli, the other processing. One is well-modeled; the other was the reason for the session.
- The rods and cones of the eye are not located along a grid. It is possible to use normalized convolution to produce acceptable images from a random sample of as few as 5% of the pixels on a regular grid, approximating the visual field. It’s possible to do even better by mimicking the distribution of cones, which are most dense around the fovea.
- The human visual system performs a form of edge detection in the visual cortex. The principles of these saliency maps can be applied to digital image processing. (For example: Plinio and Li Zhaoping.)
Also of interest: Hiroshi Momiji’s Retinal Vision for Engineers.
* — Perhaps I’ve been out of academia too long. Perhaps medical imaging does less for me than in the past. Perhaps it was just an off-year at SPIE Medical. At any rate, I didn’t attend many paper sessions, but those I did hear were a little disappointing.