Conférence

Opponency revisited

Réalisation : 10 mars 2022 Mise en ligne : 18 mars 2022
  • document 1 document 2 document 3
  • niveau 1 niveau 2 niveau 3
  • audio 1 audio 2 audio 3
Descriptif

According to the efficient coding hypothesis, the goal of the visual system should be to encode the information presented to the retina with as little redundancy as possible. From a signal processing point of view, the first step in  removing redundancy is de-correlation, which removes the second order dependencies in the signal. This principle was explored in the context of trichromatic vision by Buchsbaum and Gottschalk (1) and later Ruderman et al. (2) who found that linear de-correlation of the LMS cone responses matches the opponent color coding in the human visual system.

And yet, there is comparatively little research in image processing and computer vision that explicitly model and incorporate color opponency into solving imaging tasks. A common perception is that “colors” are redundant and/or too correlated to be of any interest, or that they are too complex to deal with. Within deep learning frameworks, color features are rarely considered.

In this talk, I will illustrate with several examples of our research, such as on saliency and super-pixels, that considering opponent colors can significantly improve image processing and computer vision tasks not only in image enhancement but also image segmentation, image ranking, etc. We have in addition extended the concept of “color opponency” to include near-infrared. And we show that these de-correlation concepts also apply to deep learning models in rather interesting ways.

 

(1) Trichromacy, opponent colours coding and optimum colour information transmission in the retina (Gershon Buchsbaum and A. Gottschalk)

(2) Statistics of cone responses to natural images: implications for visual coding (Daniel L. Ruderman, Thomas W. Cronin, and Chuan-Chin Chiao)

Intervenants
Thèmes
Notice
Contacter

Sur le même thème