Research Seminars

Adaptive Techniques for Image Enhancement: Expected Color Rendition and Data Fusion
Prof. Sebastiano Battiato, University of Catania
Mon 6 Sept 2004

Abstract

The wide diffusion of imaging consumer devices (Digital Still Cameras, Imaging Phones, ...) coupled with the increased overall performances capabilities, suggests color image processes aimed to perform image enhancement global or semantic based. While most enhancement techniques are completely blind to scene content, the first presented technique aims to improve the visual quality of natural Scene images by strongly relying on actual, and expected, image appearance. It is based on a two steps process: a scene classifier aimed to label each pixel as belonging to a particular chromatic class, followed by an automatic color correction step with dynamic range and intensity level preserving capabilities.

After that a simple algorithm able to manage real world scenes containing a wide range of light variations is presented. The human eye is capable to perceive contrast changes under a large variety of illumination conditions, thanks to its threshold adaptation property. This means the ability of the human visual system to perceive light intensities within wide dynamic ranges (the ratio of the highest and the lowest light value). Every time a real scene is observed the human brain performs some kind of selection on the perceived brightness values. Unfortunately there are rigorous limitations on the amount of dynamic range that common CCD/CMOS sensors are capable to acquire: under critical dynamic range of the incoming scene, information loss will be unavoidable in low or highlight portions of the range. Differing from classic high dynamic range extension techniques ourtechnique is conceptually very simple: it works on pixel basis without building extended range representations (often referred as radiancemaps) but obtaining a final image, containing both lowlight and highlight details, through a seamless weighted average of the input frames.