Wednesday, December 12, 2012

Learning ISP Self-Adapts to Various Color Filter Patterns

Stanford University applied for a patent US20120307116 "Learning of image processing pipeline for digital imaging devices" by Steven Lansel and Brian Wandell.

Abstract: "A learning technique is provided that learns how to process images by exploiting the spatial and spectral correlations inherent in image data to process and enhance images. Using a training set of input and desired output images, regression coefficients are learned that are optimal for a predefined estimation function that estimates the values at a pixel of the desired output image using a collection of similarly located pixels in the input image. Application of the learned regression coefficients is fast, robust to noise, adapts to the particulars of a dataset, and generalizes to a large variety of applications. The invention enables the use of image sensors with novel color filter array designs that offer expanded capabilities beyond existing sensors and take advantage of typical high pixel counts."

1 comment:

  1. Very nice.

    It would be interesting to see more context aware algorithms in the ISP pipeline, even for noise reduction, exposure or feature extraction.

    One would have to say though that, the color filter is a product of the sensor manufacture and hence known. This algorithm would be useful in prototyping non Bayer type grids, but a dedicated ISP would probably be better once the specific filter is known and the ISP algorithm optimised for that pattern.


All comments are moderated to avoid spam and personal attacks.