| Summary: | 碩士 === 國立臺灣科技大學 === 資訊工程系 === 105 === Given an input Bayer mosaic image, demosaicing is a necessary process to reconstruct the red-green-blue (RGB) full-color image. However, noise is usually unavoidable in the Bayer mosaic image, and it often degrades the quality of the reconstructed RGB full-color image. In this thesis, a novel and efficient four-stage denoising-first demosaicing method is proposed to demosaic the noisy Bayer mosaic image, leading to good quality of the reconstructed RGB full-color image. In the first stage, a texture-consistency based true noise determination approach is proposed to distinguish the true noise from the pseudo-noise in the noisy Bayer mosaic image. In the second stage, first we propose a priority-first- and edge-sensing-based recovery approach to estimate the truly noisy ground-truth G pixels. Then an interpolated R/B field-guided recovery approach is proposed to estimate the missing G pixels, for which the co-located noise-free R/B pixels are available. The remaining missing G pixels are estimated by the proposed gradient orientation-based fusion approach. In the third stage, a reconstructed G color plane-guided approach is proposed to estimate the truly noisy ground-truth R/B pixels, and an upper bound of the allowable noise level is provided to guarantee the complete recovery of the noisy Bayer mosaic image. In the fourth stage, an existing demosaicing method is adopted to reconstruct the RGB full-color image. Based on two typical noisy Bayer mosaic test image sets, which are generated from the IMAX and Kodak test image sets, the experimental results demonstrated that the proposed denoising-first demosaicing method achieves substantial quality improvement of the reconstructed Bayer mosaic and RGB full-color images when compared with several state-of-the-art comparative methods.
|
|---|
