IMAGE PROCESSING: Image de-hazing technique uses skyless calibration
Researchers from the Technion-Israel Institute of Technology (Haifa, Israel) have developed an image de-hazing technique that estimates atmospheric polarization and haze-subtraction factors based on the imaged scene, even without having a portion of the sky visible in the image.
Researchers from the Technion-Israel Institute of Technology (Haifa, Israel) have developed an image de-hazing technique that estimates atmospheric polarization and haze-subtraction factors based on the imaged scene, even without having a portion of the sky visible in the image.1 Algorithms for reducing image haze depend on certain parameters of the airlight (light-path radiance) such as the degree of polarization (DOP). These parameters can be calibrated rather easily when a portion of the sky is visible in the image. Understanding how to process images with a minimum of available information is important to other imaging applications in underwater scenes, for biological tissue analysis, and for defense and security surveillance.
A basic understanding of the mathematical specifics of a hazy image is important before understanding how those images can be de-hazed. Then, the de-hazing process inverts the image-formation equations back to the basic de-hazed irradiance value. Basically, an acquired image frame is the combination of two major components: object radiance from a scene in a clear atmosphere, and line-of-sight airlight caused by scattering of ambient light–a major cause of signal contrast reduction. The equations of the physical model depend on a number of fixed parameters such as the DOP and the haze radiance in the absence of objects.
Estimation of these parameters is essential for de-hazing. For polarization-based de-hazing algorithms, the parameters such as DOP have been estimated from image pixels in the sky near the horizon. Unfortunately, sky is not always present in an image of interest. So to define a “skyless” de-hazing algorithm, the research team utilized several basic assumptions in its work, including the observation that primary scattering of light from objects dominates secondary scattering effects, which cause blur and depolarization, and that the distribution of scattering particles in a scene is spatially homogeneous (although this approximation causes small but noticeable image deviations).
Using distance and feature cues
The two primary methods (secondary methods are also detailed in the reference) for image de-hazing developed by the researchers rely on distance-based cues and feature-based cues (see figure). In the distance-based method, the model parameters are estimated based on known distances to similar objects or portions of scenes in the image field of view. The knowledge that these two scene portions or objects have similar radiance in the absence of light scattering allows the calculation of the DOP and consequently the de-hazed radiance. Use of independent component-analysis algorithms can deliver results with more relaxed knowledge about the scene. Specifically, the distance-based method can rely on a couple of objects at known relative distances, but not necessarily having similar reflectance. On the other hand, in a feature-based method, the parameter estimation relies on two similar objects, with no knowledge about their distances.
“The skyless de-hazing technology is patented by the Technion, and we seek commercialization possibilities,” says researcher Yoav Schechner. “Besides de-hazing, we hypothesize that the results may be beneficial to atmospheric studies. The model leads to rather simple extraction of atmospheric parameters in field measurements, along very long paths, requiring no special hardware beyond a camera and a polarizer. The extracted parameters may potentially help in fitting models of atmospheric contents in the wide-ranging scene.”
- E. Namer et al., Optics Exp. 17 (2) p. 472, (Jan. 19, 2009)