Simultaneous Illumination Normalization and Sparse Image Compression via Unfolded Recovery

OpenReview Anonymous Preprint Submission787 Authors

Keywords: Signal Processing, Compressive Sensing

Abstract:

The traditional camera pipeline samples at Shannon–Nyquist rates, compresses to a sparse representation, and only afterwards performs illumination correction on the decoded image. We instead solve a single inverse problem in (reflectance, illumination) jointly, replacing the conventional CS–then–ISP order with one block-coordinate optimization. The joint pipeline beats sequential by up to 4.7 dB on textured scenes, and we report a clean failure mode on near-uniform scenes that explicitly falsifies the
identifiability gate one would naturally propose.

Submission Number: 787