Object Removal from Still Images employing Inpainting Techniques

Organisatorisches

• Art der Arbeit: Diplomarbeit
• Status: abgeschlossen
• Interne Betreuer:
  • Peter Decker
  • Dietrich Paulus
• Student:
  • Sebastian Vetter
• Beginn: 15.11.2008
• Ende: 15.05.2009

Beschreibung

When working with images taken at different light conditions or with different exposure, identifying a certain colour throughout a set of images can become a non-trivial task. One way to achieve a certain level of confidence is by placing a colour checker chart somewhere within each frame before taking the pictures and then using it for colour calibration and further colour related processing. Although this makes it much easier to work with colours in multiple images, the chart being visible in the image is not desired when displayed.

Ideally, the chart should be removed from each image before it is displayed and the missing pixel values should be replaced with the most suitable values computed from the information available in the frame. One of the more recent fields of research under the name of Image Inpainting generated several approaches to address this problem.

Estimating the missing pixel values based on data that can be extracted from an image in the spatial domain is proposed in a patch-based inpainting approach by Criminisi et al. [Criminisi2004RFA]. They find a most suitable block in the image for each block that is to be filled and try to propagate textures (linear structures) by employing it in a specific order.

Other approaches try to exploit the properties of an image in the frequency domain to generate the missing data. The Fourier domain as a pure frequency domain has been considered as well as the time-frequency domain of the Wavelet Transform. Whereas a method proposed by Aach et al. [Aach2001DII] calculates the missing values only in the frequency domain, extrapolating the most dominant frequency iteratively, Ignacio et al. [Ignacio2007BII] work in the Wavelet domain estimataing wavelet coefficients in order of specific priorities that are similar to the ones used in [Criminisi2004RFA].

This work aims at finding the most suitable algorithm for the problem described above. Therefore, an application will be developed that implements all three algorithms in their proposed outline and allows evaluation of the results. These will be compared in two ways, qualitatively and quantitatively, where the first is considered to be good visual perception of the in-painted image and the latter measures the peak signal-to-noise ration (PSNR) of the original image and the inpainted result.

Evaluating these results will uncover drawbacks and advantages of each individual approach from which a set of improvements and extensions derive. In combination with the gained experience, an algorithm is assembled that provides best result for removing the colour checker chart from an image by simply defining its position within the image.