Non-parametric Methods for Automatic Exposure Control, Radiometric Calibration and Dynamic Range Compression

Abstract

Imaging systems are essential to a wide range of modern day applications. With the continuous advancement in imaging systems, there is an on-going need to adapt and improve the imaging pipeline running inside the imaging systems.

In this thesis, methods are presented to improve the imaging pipeline of digital cameras. Here we present three methods to improve important phases of the imaging process, which are (i) ``Automatic exposure adjustment'' (ii) ``Radiometric calibration'' (iii) ''High dynamic range compression''. These contributions touch the initial, intermediate and final stages of imaging pipeline of digital cameras.

For exposure control, we propose two methods. The first makes use of CCD-based equations to formulate the exposure control problem. To estimate the exposure time, an initial image was acquired for each wavelength channel to which contrast adjustment techniques were applied. This helps to recover a reference cumulative distribution function of image brightness at each channel. The second method proposed for automatic exposure control is an iterative method applicable for a broad range of imaging systems. It uses spectral sensitivity functions such as the photopic response functions for the generation of a spectral power image of the captured scene. A target image is then generated using the spectral power image by applying histogram equalization. The exposure time is hence calculated iteratively by minimizing the squared difference between target and the current spectral power image. Here we further analyze the method by performing its stability and controllability analysis using a state space representation used in control theory. The applicability of the proposed method for exposure time calculation was shown on real world scenes using cameras with varying architectures. Radiometric calibration is the estimate of the non-linear mapping of the input radiance map to the output brightness values. The radiometric mapping is represented by the camera response function with which the radiance map of the scene is estimated. Our radiometric calibration method employs an L1 cost function by taking advantage of Weisfeld optimization scheme. The proposed calibration works with multiple input images of the scene with varying exposure. It can also perform calibration using a single input with few constraints. The proposed method outperforms, quantitatively and qualitatively, various alternative methods found in the literature of radiometric calibration.

Finally, to realistically represent the estimated radiance maps on low dynamic range display (LDR) devices, we propose a method for dynamic range compression. Radiance maps generally have higher dynamic range (HDR) as compared to the widely used display devices. Thus, for display purposes, dynamic range compression is required on HDR images. Our proposed method generates few LDR images from the HDR radiance map by clipping its values at different exposures. Using contrast information of each LDR image generated, the method uses an energy minimization approach to estimate the probability map of each LDR image. These probability maps are then used as label set to form final compressed dynamic range image for the display device. The results of our method were compared qualitatively and quantitatively with those produced by widely cited and professionally used methods.