Both in stereoscopy and in non-local image denoising a measure is required to determine to what an extent two image patches are similar despite noise and natural image variations.In this course two statistical approaches were presented to compare image patches:
The first one [1] is motivated by stereo applications and is based on a purely a contrario model. We only keep matches having a very low probability \(p_0\) to occur randomly under the background model (\(H_0\) hypothesis) that assumes that both patches are different. In this case building a suitable background model for natural images is essential.
The second one [2] is motivated by denoising applications [3] where noise levels are high or non-gaussian. This approach computes the probability \(p_1\) that two patches ressemble as much as the observations, under the \(H_1\) hypothesis that they are two independent observations of the same image only differing by noise. This probability is compared to \(p_0\) under a very simplistic background model.
This project aims to study the complementarity of both approaches for denoising and stereo applications. Time permitting a synthesis of both approaches will be studied, in the form of a novel patch distance that takes both noise and prior information into account. Testing of the novel metric can be done using the framework in [2] or directly for non-local SAR image denoising, optical image denoising, or stereo.
Programming
Matlab & C++ (optional)
Supervision
A. Almansa & F. Tupin
References
[1] N. Sabater, A. Almansa and J.-M. Morel, Meaningful Matches in Stereovision, IEEE Transactions on Pattern Analysis and Machine Intelligence, October 2011, vol. 99, [PDF] [HAL] [online demo] [source].
[2] C-A. Deledalle, L. Denis, and F. Tupin. “How to Compare Noisy Patches? Patch Similarity Beyond Gaussian Noise.” International Journal of Computer Vision (March 1): 86–102, 2012. [doi], [PDF], [HAL].
[3] Deledalle, C.-A., Tupin, F., & Denis, L. (2010). Poisson NL means: Unsupervised non local means for poisson noise. In 2010 IEEE International Conference on Image Processing. 345 E 47TH ST, New York, NY 10017 USA: IEEE.